├── README.md
├── build
    └── CMakeLists.txt
├── include
    ├── deform_node.h
    ├── dual_quat.h
    ├── dynamic_fusion.h
    ├── frame.h
    ├── nanoflann
    │   └── include
    │   │   └── nanoflann.hpp
    ├── point.h
    └── warp_field.h
├── installDependencies.sh
├── projectBuild.sh
├── report
    ├── build.sh
    ├── report.bib
    └── report.tex
└── src
    ├── deform_node.cpp
    ├── dual_quat.cpp
    ├── dynamic_fusion.cpp
    ├── main.cpp
    ├── point.cpp
    ├── saxpy.cu
    └── warp_field.cpp


/README.md:
--------------------------------------------------------------------------------
1 | # dynamicfusion
2 | BEng Individual project - DynamicFusion: Reconstruction and Tracking of Non-rigid Scenes in Real-Time
3 | 


--------------------------------------------------------------------------------
/build/CMakeLists.txt:
--------------------------------------------------------------------------------
 1 | cmake_minimum_required(VERSION 3.5)
 2 | set (CMAKE_CXX_STANDARD 11)
 3 | 
 4 | project(dynamicfusion)
 5 | 
 6 | find_package(Ceres REQUIRED)
 7 | 
 8 | set(INCLUDE_DIR
 9 |     ../include
10 |     ../include/nanoflann/include
11 |     ${CERES_INCLUDE_DIRS}
12 | )
13 | 
14 | include_directories(${INCLUDE_DIR})
15 | 
16 | set(SOURCES
17 |     ../src/main.cpp
18 | 
19 |     ../include/point.h
20 |     ../src/point.cpp
21 | 
22 |     ../include/dual_quat.h
23 |     ../src/dual_quat.cpp
24 | 
25 |     ../include/dynamic_fusion.h
26 |     ../src/dynamic_fusion.cpp
27 | 
28 |     ../include/deform_node.h
29 |     ../src/deform_node.cpp
30 | 
31 |     ../include/frame.h
32 | 
33 |     ../include/warp_field.h
34 |     ../src/warp_field.cpp
35 | 
36 |     ../include/nanoflann/include/nanoflann.hpp
37 | )
38 | 
39 | add_executable(dynamicfusion ${SOURCES})
40 | target_link_libraries(dynamicfusion ${CERES_LIBRARIES})
41 | 
42 | 


--------------------------------------------------------------------------------
/include/deform_node.h:
--------------------------------------------------------------------------------
 1 | #ifndef DEFORM_NODE_H
 2 | #define DEFORM_NODE_H
 3 | 
 4 | #include "point.h"
 5 | #include "dual_quat.h"
 6 | 
 7 | class warp_field_t;
 8 | 
 9 | /*
10 |  *  a class representing one of the deformation nodes that make up the 
11 |  *  warp field.
12 |  */
13 | class deform_node_t {
14 | public:
15 |     // constructor
16 |     deform_node_t(warp_field_t * wf, point_t p, dual_quat_t dq, float w);
17 | 
18 |     // energy function used for enforcing warp field regularisation
19 |     float regularisation(float phi);
20 |     
21 |     // weight for altering radius of influence
22 |     float weight(point_t p);
23 | 
24 |     // getters
25 |     dual_quat_t get_transform();
26 |     point_t get_position();
27 |     
28 | private:
29 |     // pointer to warp field to allow node to find neighbours
30 |     warp_field_t * warp_field;
31 | 
32 |     // properties of deformation node as in dynamic fusion paper
33 |     point_t position;
34 |     dual_quat_t transform;
35 |     float radial_weight;
36 | };
37 | 
38 | #endif
39 | 


--------------------------------------------------------------------------------
/include/dual_quat.h:
--------------------------------------------------------------------------------
 1 | #ifndef DUAL_QUAT_H
 2 | #define DUAL_QUAT_H
 3 | 
 4 | #include <boost/math/quaternion.hpp>
 5 | 
 6 | typedef boost::math::quaternion<float> quat_t;
 7 | 
 8 | class dual_quat_t {
 9 | public:
10 |     /*
11 |      *  constructors
12 |      */
13 |     dual_quat_t();
14 |     dual_quat_t(quat_t real, quat_t dual);
15 | 
16 |     /*
17 |      *  overloaded operators
18 |      */ 
19 |     dual_quat_t operator*(float scale);
20 |     void operator+=(const dual_quat_t & other);
21 | 
22 |     /*
23 |      *  public methods
24 |      */    
25 |     void normalise();
26 | 
27 | private:
28 |     // Rotational part 
29 |     quat_t real;
30 |     
31 |     // translation part
32 |     quat_t dual;
33 | };
34 | #endif
35 | 


--------------------------------------------------------------------------------
/include/dynamic_fusion.h:
--------------------------------------------------------------------------------
 1 | #ifndef DYNAMIC_FUSION_H
 2 | #define DYNAMIC_FUSION_H
 3 | 
 4 | #include "warp_field.h"
 5 | 
 6 | class dynamic_fusion_t {
 7 | public:
 8 |     /*
 9 |       constructors and destructors
10 |     */
11 |     dynamic_fusion_t();
12 |     ~dynamic_fusion_t();
13 | 
14 |     // main public method
15 |     void execute();
16 | 
17 | private:
18 |     warp_field_t warp_field;
19 | };
20 | 
21 | #endif
22 | 


--------------------------------------------------------------------------------
/include/frame.h:
--------------------------------------------------------------------------------
 1 | #ifndef FRAME_H
 2 | #define FRAME_H
 3 | 
 4 | #include "point.h"
 5 | 
 6 | #include <vector>
 7 | 
 8 | /*
 9 |     a small container class representing a frame.
10 |     contains a unique id, and the vertex positions and normals of the frame
11 | */
12 | class frame_t {
13 | public:
14 |     // constructor
15 |     frame_t(int frame_id);
16 | 
17 |     // destructor
18 |     ~frame_t();
19 | 
20 | private:
21 |     // unique id for frame
22 |     int frame_id;
23 | 
24 |     // point clouds storing position and normal data
25 |     std::vector<point_t> vertices;
26 |     std::vector<point_t> normals;
27 | };
28 | 
29 | #endif
30 | 


--------------------------------------------------------------------------------
/include/nanoflann/include/nanoflann.hpp:
--------------------------------------------------------------------------------
   1 | /***********************************************************************
   2 |  * Software License Agreement (BSD License)
   3 |  *
   4 |  * Copyright 2008-2009  Marius Muja (mariusm@cs.ubc.ca). All rights reserved.
   5 |  * Copyright 2008-2009  David G. Lowe (lowe@cs.ubc.ca). All rights reserved.
   6 |  * Copyright 2011-2016  Jose Luis Blanco (joseluisblancoc@gmail.com).
   7 |  *   All rights reserved.
   8 |  *
   9 |  * THE BSD LICENSE
  10 |  *
  11 |  * Redistribution and use in source and binary forms, with or without
  12 |  * modification, are permitted provided that the following conditions
  13 |  * are met:
  14 |  *
  15 |  * 1. Redistributions of source code must retain the above copyright
  16 |  *    notice, this list of conditions and the following disclaimer.
  17 |  * 2. Redistributions in binary form must reproduce the above copyright
  18 |  *    notice, this list of conditions and the following disclaimer in the
  19 |  *    documentation and/or other materials provided with the distribution.
  20 |  *
  21 |  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  22 |  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  23 |  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  24 |  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
  25 |  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
  26 |  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  27 |  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  28 |  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  29 |  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
  30 |  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  31 |  *************************************************************************/
  32 | 
  33 | /** \mainpage nanoflann C++ API documentation
  34 |   *  nanoflann is a C++ header-only library for building KD-Trees, mostly
  35 |   *  optimized for 2D or 3D point clouds.
  36 |   *
  37 |   *  nanoflann does not require compiling or installing, just an
  38 |   *  #include <nanoflann.hpp> in your code.
  39 |   *
  40 |   *  See:
  41 |   *   - <a href="modules.html" >C++ API organized by modules</a>
  42 |   *   - <a href="https://github.com/jlblancoc/nanoflann" >Online README</a>
  43 |   *   - <a href="http://jlblancoc.github.io/nanoflann/" >Doxygen documentation</a>
  44 |   */
  45 | 
  46 | #ifndef  NANOFLANN_HPP_
  47 | #define  NANOFLANN_HPP_
  48 | 
  49 | #include <vector>
  50 | #include <cassert>
  51 | #include <algorithm>
  52 | #include <stdexcept>
  53 | #include <cstdio>  // for fwrite()
  54 | #define _USE_MATH_DEFINES // Required by MSVC to define M_PI,etc. in <cmath>
  55 | #include <cmath>   // for abs()
  56 | #include <cstdlib> // for abs()
  57 | #include <limits>
  58 | #include <math.h>
  59 | 
  60 | // Avoid conflicting declaration of min/max macros in windows headers
  61 | #if !defined(NOMINMAX) && (defined(_WIN32) || defined(_WIN32_)  || defined(WIN32) || defined(_WIN64))
  62 | # define NOMINMAX
  63 | # ifdef max
  64 | #  undef   max
  65 | #  undef   min
  66 | # endif
  67 | #endif
  68 | 
  69 | namespace nanoflann
  70 | {
  71 | /** @addtogroup nanoflann_grp nanoflann C++ library for ANN
  72 |   *  @{ */
  73 | 
  74 |   	/** Library version: 0xMmP (M=Major,m=minor,P=patch) */
  75 | 	#define NANOFLANN_VERSION 0x123
  76 | 
  77 | 	/** @addtogroup result_sets_grp Result set classes
  78 | 	  *  @{ */
  79 | 	template <typename DistanceType, typename IndexType = size_t, typename CountType = size_t>
  80 | 	class KNNResultSet
  81 | 	{
  82 | 		IndexType * indices;
  83 | 		DistanceType* dists;
  84 | 		CountType capacity;
  85 | 		CountType count;
  86 | 
  87 | 	public:
  88 | 		inline KNNResultSet(CountType capacity_) : indices(0), dists(0), capacity(capacity_), count(0)
  89 | 		{
  90 | 		}
  91 | 
  92 | 		inline void init(IndexType* indices_, DistanceType* dists_)
  93 | 		{
  94 | 			indices = indices_;
  95 | 			dists = dists_;
  96 | 			count = 0;
  97 |             if (capacity)
  98 |                 dists[capacity-1] = (std::numeric_limits<DistanceType>::max)();
  99 | 		}
 100 | 
 101 | 		inline CountType size() const
 102 | 		{
 103 | 			return count;
 104 | 		}
 105 | 
 106 | 		inline bool full() const
 107 | 		{
 108 | 			return count == capacity;
 109 | 		}
 110 | 
 111 | 
 112 |                 /**
 113 |                  * Called during search to add an element matching the criteria.
 114 |                  * @return true if the search should be continued, false if the results are sufficient
 115 |                  */
 116 |                 inline bool addPoint(DistanceType dist, IndexType index)
 117 | 		{
 118 | 			CountType i;
 119 | 			for (i = count; i > 0; --i) {
 120 | #ifdef NANOFLANN_FIRST_MATCH   // If defined and two points have the same distance, the one with the lowest-index will be returned first.
 121 | 				if ( (dists[i-1] > dist) || ((dist == dists[i-1]) && (indices[i-1] > index)) ) {
 122 | #else
 123 | 				if (dists[i-1] > dist) {
 124 | #endif
 125 | 					if (i < capacity) {
 126 | 						dists[i] = dists[i-1];
 127 | 						indices[i] = indices[i-1];
 128 | 					}
 129 | 				}
 130 | 				else break;
 131 | 			}
 132 | 			if (i < capacity) {
 133 | 				dists[i] = dist;
 134 | 				indices[i] = index;
 135 | 			}
 136 | 			if (count < capacity) count++;
 137 | 
 138 |                         // tell caller that the search shall continue
 139 |                         return true;
 140 | 		}
 141 | 
 142 | 		inline DistanceType worstDist() const
 143 | 		{
 144 | 			return dists[capacity-1];
 145 | 		}
 146 | 	};
 147 | 
 148 | 	/** operator "<" for std::sort() */
 149 | 	struct IndexDist_Sorter
 150 | 	{
 151 | 		/** PairType will be typically: std::pair<IndexType,DistanceType> */
 152 | 		template <typename PairType>
 153 | 		inline bool operator()(const PairType &p1, const PairType &p2) const {
 154 | 			return p1.second < p2.second;
 155 | 		}
 156 | 	};
 157 | 
 158 | 	/**
 159 | 	 * A result-set class used when performing a radius based search.
 160 | 	 */
 161 | 	template <typename DistanceType, typename IndexType = size_t>
 162 | 	class RadiusResultSet
 163 | 	{
 164 | 	public:
 165 | 		const DistanceType radius;
 166 | 
 167 | 		std::vector<std::pair<IndexType, DistanceType> > &m_indices_dists;
 168 | 
 169 | 		inline RadiusResultSet(DistanceType radius_, std::vector<std::pair<IndexType,DistanceType> > &indices_dists) : radius(radius_), m_indices_dists(indices_dists)
 170 | 		{
 171 | 			init();
 172 | 		}
 173 | 
 174 | 		inline void init() { clear(); }
 175 | 		inline void clear() { m_indices_dists.clear(); }
 176 | 
 177 | 		inline size_t size() const { return m_indices_dists.size(); }
 178 | 
 179 | 		inline bool full() const { return true; }
 180 | 
 181 |                 /**
 182 |                  * Called during search to add an element matching the criteria.
 183 |                  * @return true if the search should be continued, false if the results are sufficient
 184 |                  */
 185 |                 inline bool addPoint(DistanceType dist, IndexType index)
 186 | 		{
 187 | 			if (dist < radius)
 188 | 				m_indices_dists.push_back(std::make_pair(index, dist));
 189 |                         return true;
 190 | 		}
 191 | 
 192 | 		inline DistanceType worstDist() const { return radius; }
 193 | 
 194 | 		/**
 195 | 		 * Find the worst result (furtherest neighbor) without copying or sorting
 196 | 		 * Pre-conditions: size() > 0
 197 | 		 */
 198 | 		std::pair<IndexType,DistanceType> worst_item() const
 199 | 		{
 200 | 		   if (m_indices_dists.empty()) throw std::runtime_error("Cannot invoke RadiusResultSet::worst_item() on an empty list of results.");
 201 | 		   typedef typename std::vector<std::pair<IndexType, DistanceType> >::const_iterator DistIt;
 202 | 		   DistIt it = std::max_element(m_indices_dists.begin(), m_indices_dists.end(), IndexDist_Sorter());
 203 | 		   return *it;
 204 | 		}
 205 | 	};
 206 | 
 207 | 
 208 | 	/** @} */
 209 | 
 210 | 
 211 | 	/** @addtogroup loadsave_grp Load/save auxiliary functions
 212 | 	  * @{ */
 213 | 	template<typename T>
 214 | 	void save_value(FILE* stream, const T& value, size_t count = 1)
 215 | 	{
 216 | 		fwrite(&value, sizeof(value), count, stream);
 217 | 	}
 218 | 
 219 | 	template<typename T>
 220 | 	void save_value(FILE* stream, const std::vector<T>& value)
 221 | 	{
 222 | 		size_t size = value.size();
 223 | 		fwrite(&size, sizeof(size_t), 1, stream);
 224 | 		fwrite(&value[0], sizeof(T), size, stream);
 225 | 	}
 226 | 
 227 | 	template<typename T>
 228 | 	void load_value(FILE* stream, T& value, size_t count = 1)
 229 | 	{
 230 | 		size_t read_cnt = fread(&value, sizeof(value), count, stream);
 231 | 		if (read_cnt != count) {
 232 | 			throw std::runtime_error("Cannot read from file");
 233 | 		}
 234 | 	}
 235 | 
 236 | 
 237 | 	template<typename T>
 238 | 	void load_value(FILE* stream, std::vector<T>& value)
 239 | 	{
 240 | 		size_t size;
 241 | 		size_t read_cnt = fread(&size, sizeof(size_t), 1, stream);
 242 | 		if (read_cnt != 1) {
 243 | 			throw std::runtime_error("Cannot read from file");
 244 | 		}
 245 | 		value.resize(size);
 246 | 		read_cnt = fread(&value[0], sizeof(T), size, stream);
 247 | 		if (read_cnt != size) {
 248 | 			throw std::runtime_error("Cannot read from file");
 249 | 		}
 250 | 	}
 251 | 	/** @} */
 252 | 
 253 | 
 254 | 	/** @addtogroup metric_grp Metric (distance) classes
 255 | 	  * @{ */
 256 | 
 257 | 	struct Metric
 258 | 	{
 259 | 	};
 260 | 
 261 | 	/** Manhattan distance functor (generic version, optimized for high-dimensionality data sets).
 262 | 	  *  Corresponding distance traits: nanoflann::metric_L1
 263 | 	  * \tparam T Type of the elements (e.g. double, float, uint8_t)
 264 | 	  * \tparam _DistanceType Type of distance variables (must be signed) (e.g. float, double, int64_t)
 265 | 	  */
 266 | 	template<class T, class DataSource, typename _DistanceType = T>
 267 | 	struct L1_Adaptor
 268 | 	{
 269 | 		typedef T ElementType;
 270 | 		typedef _DistanceType DistanceType;
 271 | 
 272 | 		const DataSource &data_source;
 273 | 
 274 | 		L1_Adaptor(const DataSource &_data_source) : data_source(_data_source) { }
 275 | 
 276 | 		inline DistanceType evalMetric(const T* a, const size_t b_idx, size_t size, DistanceType worst_dist = -1) const
 277 | 		{
 278 | 			DistanceType result = DistanceType();
 279 | 			const T* last = a + size;
 280 | 			const T* lastgroup = last - 3;
 281 | 			size_t d = 0;
 282 | 
 283 | 			/* Process 4 items with each loop for efficiency. */
 284 | 			while (a < lastgroup) {
 285 | 				const DistanceType diff0 = std::abs(a[0] - data_source.kdtree_get_pt(b_idx,d++));
 286 | 				const DistanceType diff1 = std::abs(a[1] - data_source.kdtree_get_pt(b_idx,d++));
 287 | 				const DistanceType diff2 = std::abs(a[2] - data_source.kdtree_get_pt(b_idx,d++));
 288 | 				const DistanceType diff3 = std::abs(a[3] - data_source.kdtree_get_pt(b_idx,d++));
 289 | 				result += diff0 + diff1 + diff2 + diff3;
 290 | 				a += 4;
 291 | 				if ((worst_dist > 0) && (result > worst_dist)) {
 292 | 					return result;
 293 | 				}
 294 | 			}
 295 | 			/* Process last 0-3 components.  Not needed for standard vector lengths. */
 296 | 			while (a < last) {
 297 | 				result += std::abs( *a++ - data_source.kdtree_get_pt(b_idx, d++) );
 298 | 			}
 299 | 			return result;
 300 | 		}
 301 | 
 302 | 		template <typename U, typename V>
 303 | 		inline DistanceType accum_dist(const U a, const V b, int ) const
 304 | 		{
 305 | 			return std::abs(a-b);
 306 | 		}
 307 | 	};
 308 | 
 309 | 	/** Squared Euclidean distance functor (generic version, optimized for high-dimensionality data sets).
 310 | 	  *  Corresponding distance traits: nanoflann::metric_L2
 311 | 	  * \tparam T Type of the elements (e.g. double, float, uint8_t)
 312 | 	  * \tparam _DistanceType Type of distance variables (must be signed) (e.g. float, double, int64_t)
 313 | 	  */
 314 | 	template<class T, class DataSource, typename _DistanceType = T>
 315 | 	struct L2_Adaptor
 316 | 	{
 317 | 		typedef T ElementType;
 318 | 		typedef _DistanceType DistanceType;
 319 | 
 320 | 		const DataSource &data_source;
 321 | 
 322 | 		L2_Adaptor(const DataSource &_data_source) : data_source(_data_source) { }
 323 | 
 324 | 		inline DistanceType evalMetric(const T* a, const size_t b_idx, size_t size, DistanceType worst_dist = -1) const
 325 | 		{
 326 | 			DistanceType result = DistanceType();
 327 | 			const T* last = a + size;
 328 | 			const T* lastgroup = last - 3;
 329 | 			size_t d = 0;
 330 | 
 331 | 			/* Process 4 items with each loop for efficiency. */
 332 | 			while (a < lastgroup) {
 333 | 				const DistanceType diff0 = a[0] - data_source.kdtree_get_pt(b_idx,d++);
 334 | 				const DistanceType diff1 = a[1] - data_source.kdtree_get_pt(b_idx,d++);
 335 | 				const DistanceType diff2 = a[2] - data_source.kdtree_get_pt(b_idx,d++);
 336 | 				const DistanceType diff3 = a[3] - data_source.kdtree_get_pt(b_idx,d++);
 337 | 				result += diff0 * diff0 + diff1 * diff1 + diff2 * diff2 + diff3 * diff3;
 338 | 				a += 4;
 339 | 				if ((worst_dist > 0) && (result > worst_dist)) {
 340 | 					return result;
 341 | 				}
 342 | 			}
 343 | 			/* Process last 0-3 components.  Not needed for standard vector lengths. */
 344 | 			while (a < last) {
 345 | 				const DistanceType diff0 = *a++ - data_source.kdtree_get_pt(b_idx, d++);
 346 | 				result += diff0 * diff0;
 347 | 			}
 348 | 			return result;
 349 | 		}
 350 | 
 351 | 		template <typename U, typename V>
 352 | 		inline DistanceType accum_dist(const U a, const V b, int ) const
 353 | 		{
 354 | 			return (a - b) * (a - b);
 355 | 		}
 356 | 	};
 357 | 
 358 | 	/** Squared Euclidean (L2) distance functor (suitable for low-dimensionality datasets, like 2D or 3D point clouds)
 359 | 	  *  Corresponding distance traits: nanoflann::metric_L2_Simple
 360 | 	  * \tparam T Type of the elements (e.g. double, float, uint8_t)
 361 | 	  * \tparam _DistanceType Type of distance variables (must be signed) (e.g. float, double, int64_t)
 362 | 	  */
 363 | 	template<class T, class DataSource, typename _DistanceType = T>
 364 | 	struct L2_Simple_Adaptor
 365 | 	{
 366 | 		typedef T ElementType;
 367 | 		typedef _DistanceType DistanceType;
 368 | 
 369 | 		const DataSource &data_source;
 370 | 
 371 | 		L2_Simple_Adaptor(const DataSource &_data_source) : data_source(_data_source) { }
 372 | 
 373 | 		inline DistanceType evalMetric(const T* a, const size_t b_idx, size_t size) const {
 374 | 			DistanceType result = DistanceType();
 375 | 			for (size_t i = 0; i < size; ++i) {
 376 | 				const DistanceType diff = a[i] - data_source.kdtree_get_pt(b_idx, i);
 377 | 				result += diff * diff;
 378 | 			}
 379 | 			return result;
 380 | 		}
 381 | 
 382 | 		template <typename U, typename V>
 383 | 		inline DistanceType accum_dist(const U a, const V b, int ) const
 384 | 		{
 385 | 			return (a - b) * (a - b);
 386 | 		}
 387 | 	};
 388 | 
 389 | 	/** SO2 distance functor
 390 | 	  *  Corresponding distance traits: nanoflann::metric_SO2
 391 | 	  * \tparam T Type of the elements (e.g. double, float)
 392 | 	  * \tparam _DistanceType Type of distance variables (must be signed) (e.g. float, double)
 393 | 	  * orientation is constrained to be in [-pi, pi]
 394 | 	  */
 395 | 	template<class T, class DataSource, typename _DistanceType = T>
 396 | 	struct SO2_Adaptor
 397 | 	{
 398 | 		typedef T ElementType;
 399 | 		typedef _DistanceType DistanceType;
 400 | 
 401 | 		const DataSource &data_source;
 402 | 
 403 | 		SO2_Adaptor(const DataSource &_data_source) : data_source(_data_source) { }
 404 | 
 405 | 		inline DistanceType evalMetric(const T* a, const size_t b_idx, size_t size) const {
 406 | 			return accum_dist(a[size-1], data_source.kdtree_get_pt(b_idx, size - 1) , size - 1);
 407 | 		}
 408 | 
 409 | 		template <typename U, typename V>
 410 | 		inline DistanceType accum_dist(const U a, const V b, int ) const
 411 | 		{
 412 | 			DistanceType result = DistanceType();
 413 | 			result = b - a;
 414 | 			if (result > M_PI)
 415 | 				result -= 2. * M_PI;
 416 | 			else if (result < -M_PI)
 417 | 				result += 2. * M_PI;
 418 | 			return result;
 419 | 		}
 420 | 	};
 421 | 
 422 | 	/** SO3 distance functor (Uses L2_Simple)
 423 | 	  *  Corresponding distance traits: nanoflann::metric_SO3
 424 | 	  * \tparam T Type of the elements (e.g. double, float)
 425 | 	  * \tparam _DistanceType Type of distance variables (must be signed) (e.g. float, double)
 426 | 	  */
 427 | 	template<class T, class DataSource, typename _DistanceType = T>
 428 | 	struct SO3_Adaptor
 429 | 	{
 430 | 		typedef T ElementType;
 431 | 		typedef _DistanceType DistanceType;
 432 | 
 433 | 		L2_Simple_Adaptor<T, DataSource > distance_L2_Simple;
 434 | 
 435 | 		SO3_Adaptor(const DataSource &_data_source) : distance_L2_Simple(_data_source) { }
 436 | 
 437 | 		inline DistanceType evalMetric(const T* a, const size_t b_idx, size_t size) const {
 438 | 			return distance_L2_Simple.evalMetric(a, b_idx, size);
 439 | 		}
 440 | 
 441 | 		template <typename U, typename V>
 442 | 		inline DistanceType accum_dist(const U a, const V b, int idx) const
 443 | 		{
 444 | 			return distance_L2_Simple.accum_dist(a, b, idx);
 445 | 		}
 446 | 	};
 447 | 
 448 | 	/** Metaprogramming helper traits class for the L1 (Manhattan) metric */
 449 | 	struct metric_L1 : public Metric
 450 | 	{
 451 | 		template<class T, class DataSource>
 452 | 		struct traits {
 453 | 			typedef L1_Adaptor<T, DataSource> distance_t;
 454 | 		};
 455 | 	};
 456 | 	/** Metaprogramming helper traits class for the L2 (Euclidean) metric */
 457 | 	struct metric_L2 : public Metric 
 458 | 	{
 459 | 		template<class T, class DataSource>
 460 | 		struct traits {
 461 | 			typedef L2_Adaptor<T, DataSource> distance_t;
 462 | 		};
 463 | 	};
 464 | 	/** Metaprogramming helper traits class for the L2_simple (Euclidean) metric */
 465 | 	struct metric_L2_Simple : public Metric
 466 | 	{
 467 | 		template<class T, class DataSource>
 468 | 		struct traits {
 469 | 			typedef L2_Simple_Adaptor<T, DataSource> distance_t;
 470 | 		};
 471 | 	};
 472 | 	/** Metaprogramming helper traits class for the SO3_InnerProdQuat metric */
 473 | 	struct metric_SO2 : public Metric 
 474 | 	{
 475 | 		template<class T, class DataSource>
 476 | 		struct traits {
 477 | 			typedef SO2_Adaptor<T, DataSource> distance_t;
 478 | 		};
 479 | 	};
 480 | 	/** Metaprogramming helper traits class for the SO3_InnerProdQuat metric */
 481 | 	struct metric_SO3 : public Metric
 482 | 	{
 483 | 		template<class T, class DataSource>
 484 | 		struct traits {
 485 | 			typedef SO3_Adaptor<T, DataSource> distance_t;
 486 | 		};
 487 | 	};
 488 | 
 489 | 	/** @} */
 490 | 
 491 | 	/** @addtogroup param_grp Parameter structs
 492 | 	  * @{ */
 493 | 
 494 | 	/**  Parameters (see README.md) */
 495 | 	struct KDTreeSingleIndexAdaptorParams
 496 | 	{
 497 | 		KDTreeSingleIndexAdaptorParams(size_t _leaf_max_size = 10) :
 498 | 			leaf_max_size(_leaf_max_size)
 499 | 		{}
 500 | 
 501 | 		size_t leaf_max_size;
 502 | 	};
 503 | 
 504 | 	/** Search options for KDTreeSingleIndexAdaptor::findNeighbors() */
 505 | 	struct SearchParams
 506 | 	{
 507 | 		/** Note: The first argument (checks_IGNORED_) is ignored, but kept for compatibility with the FLANN interface */
 508 | 		SearchParams(int checks_IGNORED_ = 32, float eps_ = 0, bool sorted_ = true ) :
 509 | 			checks(checks_IGNORED_), eps(eps_), sorted(sorted_) {}
 510 | 
 511 | 		int   checks;  //!< Ignored parameter (Kept for compatibility with the FLANN interface).
 512 | 		float eps;  //!< search for eps-approximate neighbours (default: 0)
 513 | 		bool sorted; //!< only for radius search, require neighbours sorted by distance (default: true)
 514 | 	};
 515 | 	/** @} */
 516 | 
 517 | 
 518 | 	/** @addtogroup memalloc_grp Memory allocation
 519 | 	  * @{ */
 520 | 
 521 | 	/**
 522 | 	 * Allocates (using C's malloc) a generic type T.
 523 | 	 *
 524 | 	 * Params:
 525 | 	 *     count = number of instances to allocate.
 526 | 	 * Returns: pointer (of type T*) to memory buffer
 527 | 	 */
 528 | 	template <typename T>
 529 | 	inline T* allocate(size_t count = 1)
 530 | 	{
 531 | 		T* mem = static_cast<T*>( ::malloc(sizeof(T)*count));
 532 | 		return mem;
 533 | 	}
 534 | 
 535 | 
 536 | 	/**
 537 | 	 * Pooled storage allocator
 538 | 	 *
 539 | 	 * The following routines allow for the efficient allocation of storage in
 540 | 	 * small chunks from a specified pool.  Rather than allowing each structure
 541 | 	 * to be freed individually, an entire pool of storage is freed at once.
 542 | 	 * This method has two advantages over just using malloc() and free().  First,
 543 | 	 * it is far more efficient for allocating small objects, as there is
 544 | 	 * no overhead for remembering all the information needed to free each
 545 | 	 * object or consolidating fragmented memory.  Second, the decision about
 546 | 	 * how long to keep an object is made at the time of allocation, and there
 547 | 	 * is no need to track down all the objects to free them.
 548 | 	 *
 549 | 	 */
 550 | 
 551 | 	const size_t     WORDSIZE = 16;
 552 | 	const size_t     BLOCKSIZE = 8192;
 553 | 
 554 | 	class PooledAllocator
 555 | 	{
 556 | 		/* We maintain memory alignment to word boundaries by requiring that all
 557 | 		    allocations be in multiples of the machine wordsize.  */
 558 | 		/* Size of machine word in bytes.  Must be power of 2. */
 559 | 		/* Minimum number of bytes requested at a time from	the system.  Must be multiple of WORDSIZE. */
 560 | 
 561 | 
 562 | 		size_t  remaining;  /* Number of bytes left in current block of storage. */
 563 | 		void*   base;     /* Pointer to base of current block of storage. */
 564 | 		void*   loc;      /* Current location in block to next allocate memory. */
 565 | 
 566 | 		void internal_init()
 567 | 		{
 568 | 			remaining = 0;
 569 | 			base = NULL;
 570 | 			usedMemory = 0;
 571 | 			wastedMemory = 0;
 572 | 		}
 573 | 
 574 | 	public:
 575 | 		size_t  usedMemory;
 576 | 		size_t  wastedMemory;
 577 | 
 578 | 		/**
 579 | 		    Default constructor. Initializes a new pool.
 580 | 		 */
 581 | 		PooledAllocator() {
 582 | 			internal_init();
 583 | 		}
 584 | 
 585 | 		/**
 586 | 		 * Destructor. Frees all the memory allocated in this pool.
 587 | 		 */
 588 | 		~PooledAllocator() {
 589 | 			free_all();
 590 | 		}
 591 | 
 592 | 		/** Frees all allocated memory chunks */
 593 | 		void free_all()
 594 | 		{
 595 | 			while (base != NULL) {
 596 | 				void *prev = *(static_cast<void**>( base)); /* Get pointer to prev block. */
 597 | 				::free(base);
 598 | 				base = prev;
 599 | 			}
 600 | 			internal_init();
 601 | 		}
 602 | 
 603 | 		/**
 604 | 		 * Returns a pointer to a piece of new memory of the given size in bytes
 605 | 		 * allocated from the pool.
 606 | 		 */
 607 | 		void* malloc(const size_t req_size)
 608 | 		{
 609 | 			/* Round size up to a multiple of wordsize.  The following expression
 610 | 			    only works for WORDSIZE that is a power of 2, by masking last bits of
 611 | 			    incremented size to zero.
 612 | 			 */
 613 | 			const size_t size = (req_size + (WORDSIZE - 1)) & ~(WORDSIZE - 1);
 614 | 
 615 | 			/* Check whether a new block must be allocated.  Note that the first word
 616 | 			    of a block is reserved for a pointer to the previous block.
 617 | 			 */
 618 | 			if (size > remaining) {
 619 | 
 620 | 				wastedMemory += remaining;
 621 | 
 622 | 				/* Allocate new storage. */
 623 | 				const size_t blocksize = (size + sizeof(void*) + (WORDSIZE - 1) > BLOCKSIZE) ?
 624 | 							size + sizeof(void*) + (WORDSIZE - 1) : BLOCKSIZE;
 625 | 
 626 | 				// use the standard C malloc to allocate memory
 627 | 				void* m = ::malloc(blocksize);
 628 | 				if (!m) {
 629 | 					fprintf(stderr, "Failed to allocate memory.\n");
 630 | 					return NULL;
 631 | 				}
 632 | 
 633 | 				/* Fill first word of new block with pointer to previous block. */
 634 | 				static_cast<void**>(m)[0] = base;
 635 | 				base = m;
 636 | 
 637 | 				size_t shift = 0;
 638 | 				//int size_t = (WORDSIZE - ( (((size_t)m) + sizeof(void*)) & (WORDSIZE-1))) & (WORDSIZE-1);
 639 | 
 640 | 				remaining = blocksize - sizeof(void*) - shift;
 641 | 				loc = (static_cast<char*>(m) + sizeof(void*) + shift);
 642 | 			}
 643 | 			void* rloc = loc;
 644 | 			loc = static_cast<char*>(loc) + size;
 645 | 			remaining -= size;
 646 | 
 647 | 			usedMemory += size;
 648 | 
 649 | 			return rloc;
 650 | 		}
 651 | 
 652 | 		/**
 653 | 		 * Allocates (using this pool) a generic type T.
 654 | 		 *
 655 | 		 * Params:
 656 | 		 *     count = number of instances to allocate.
 657 | 		 * Returns: pointer (of type T*) to memory buffer
 658 | 		 */
 659 | 		template <typename T>
 660 | 		T* allocate(const size_t count = 1)
 661 | 		{
 662 | 			T* mem = static_cast<T*>(this->malloc(sizeof(T)*count));
 663 | 			return mem;
 664 | 		}
 665 | 
 666 | 	};
 667 | 	/** @} */
 668 | 
 669 | 	/** @addtogroup nanoflann_metaprog_grp Auxiliary metaprogramming stuff
 670 | 	  * @{ */
 671 | 
 672 | 	// ----------------  CArray -------------------------
 673 | 	/** A STL container (as wrapper) for arrays of constant size defined at compile time (class imported from the MRPT project)
 674 | 	 * This code is an adapted version from Boost, modifed for its integration
 675 | 	 *	within MRPT (JLBC, Dec/2009) (Renamed array -> CArray to avoid possible potential conflicts).
 676 | 	 * See
 677 | 	 *      http://www.josuttis.com/cppcode
 678 | 	 * for details and the latest version.
 679 | 	 * See
 680 | 	 *      http://www.boost.org/libs/array for Documentation.
 681 | 	 * for documentation.
 682 | 	 *
 683 | 	 * (C) Copyright Nicolai M. Josuttis 2001.
 684 | 	 * Permission to copy, use, modify, sell and distribute this software
 685 | 	 * is granted provided this copyright notice appears in all copies.
 686 | 	 * This software is provided "as is" without express or implied
 687 | 	 * warranty, and with no claim as to its suitability for any purpose.
 688 | 	 *
 689 | 	 * 29 Jan 2004 - minor fixes (Nico Josuttis)
 690 | 	 * 04 Dec 2003 - update to synch with library TR1 (Alisdair Meredith)
 691 | 	 * 23 Aug 2002 - fix for Non-MSVC compilers combined with MSVC libraries.
 692 | 	 * 05 Aug 2001 - minor update (Nico Josuttis)
 693 | 	 * 20 Jan 2001 - STLport fix (Beman Dawes)
 694 | 	 * 29 Sep 2000 - Initial Revision (Nico Josuttis)
 695 | 	 *
 696 | 	 * Jan 30, 2004
 697 | 	 */
 698 |     template <typename T, std::size_t N>
 699 |     class CArray {
 700 |       public:
 701 |         T elems[N];    // fixed-size array of elements of type T
 702 | 
 703 |       public:
 704 |         // type definitions
 705 |         typedef T              value_type;
 706 |         typedef T*             iterator;
 707 |         typedef const T*       const_iterator;
 708 |         typedef T&             reference;
 709 |         typedef const T&       const_reference;
 710 |         typedef std::size_t    size_type;
 711 |         typedef std::ptrdiff_t difference_type;
 712 | 
 713 |         // iterator support
 714 |         inline iterator begin() { return elems; }
 715 |         inline const_iterator begin() const { return elems; }
 716 |         inline iterator end() { return elems+N; }
 717 |         inline const_iterator end() const { return elems+N; }
 718 | 
 719 |         // reverse iterator support
 720 | #if !defined(BOOST_NO_TEMPLATE_PARTIAL_SPECIALIZATION) && !defined(BOOST_MSVC_STD_ITERATOR) && !defined(BOOST_NO_STD_ITERATOR_TRAITS)
 721 |         typedef std::reverse_iterator<iterator> reverse_iterator;
 722 |         typedef std::reverse_iterator<const_iterator> const_reverse_iterator;
 723 | #elif defined(_MSC_VER) && (_MSC_VER == 1300) && defined(BOOST_DINKUMWARE_STDLIB) && (BOOST_DINKUMWARE_STDLIB == 310)
 724 |         // workaround for broken reverse_iterator in VC7
 725 |         typedef std::reverse_iterator<std::_Ptrit<value_type, difference_type, iterator,
 726 |                                       reference, iterator, reference> > reverse_iterator;
 727 |         typedef std::reverse_iterator<std::_Ptrit<value_type, difference_type, const_iterator,
 728 |                                       const_reference, iterator, reference> > const_reverse_iterator;
 729 | #else
 730 |         // workaround for broken reverse_iterator implementations
 731 |         typedef std::reverse_iterator<iterator,T> reverse_iterator;
 732 |         typedef std::reverse_iterator<const_iterator,T> const_reverse_iterator;
 733 | #endif
 734 | 
 735 |         reverse_iterator rbegin() { return reverse_iterator(end()); }
 736 |         const_reverse_iterator rbegin() const { return const_reverse_iterator(end()); }
 737 |         reverse_iterator rend() { return reverse_iterator(begin()); }
 738 |         const_reverse_iterator rend() const { return const_reverse_iterator(begin()); }
 739 |         // operator[]
 740 |         inline reference operator[](size_type i) { return elems[i]; }
 741 |         inline const_reference operator[](size_type i) const { return elems[i]; }
 742 |         // at() with range check
 743 |         reference at(size_type i) { rangecheck(i); return elems[i]; }
 744 |         const_reference at(size_type i) const { rangecheck(i); return elems[i]; }
 745 |         // front() and back()
 746 |         reference front() { return elems[0]; }
 747 |         const_reference front() const { return elems[0]; }
 748 |         reference back() { return elems[N-1]; }
 749 |         const_reference back() const { return elems[N-1]; }
 750 |         // size is constant
 751 |         static inline size_type size() { return N; }
 752 |         static bool empty() { return false; }
 753 |         static size_type max_size() { return N; }
 754 |         enum { static_size = N };
 755 | 		/** This method has no effects in this class, but raises an exception if the expected size does not match */
 756 | 		inline void resize(const size_t nElements) { if (nElements!=N) throw std::logic_error("Try to change the size of a CArray."); }
 757 |         // swap (note: linear complexity in N, constant for given instantiation)
 758 |         void swap (CArray<T,N>& y) { std::swap_ranges(begin(),end(),y.begin()); }
 759 |         // direct access to data (read-only)
 760 |         const T* data() const { return elems; }
 761 |         // use array as C array (direct read/write access to data)
 762 |         T* data() { return elems; }
 763 |         // assignment with type conversion
 764 |         template <typename T2> CArray<T,N>& operator= (const CArray<T2,N>& rhs) {
 765 |             std::copy(rhs.begin(),rhs.end(), begin());
 766 |             return *this;
 767 |         }
 768 |         // assign one value to all elements
 769 |         inline void assign (const T& value) { for (size_t i=0;i<N;i++) elems[i]=value; }
 770 |         // assign (compatible with std::vector's one) (by JLBC for MRPT)
 771 |         void assign (const size_t n, const T& value) { assert(N==n); for (size_t i=0;i<N;i++) elems[i]=value; }
 772 |       private:
 773 |         // check range (may be private because it is static)
 774 |         static void rangecheck (size_type i) { if (i >= size()) { throw std::out_of_range("CArray<>: index out of range"); } }
 775 |     }; // end of CArray
 776 | 
 777 | 	/** Used to declare fixed-size arrays when DIM>0, dynamically-allocated vectors when DIM=-1.
 778 | 	  * Fixed size version for a generic DIM:
 779 | 	  */
 780 | 	template <int DIM, typename T>
 781 | 	struct array_or_vector_selector
 782 | 	{
 783 | 		typedef CArray<T, DIM> container_t;
 784 | 	};
 785 | 	/** Dynamic size version */
 786 | 	template <typename T>
 787 | 	struct array_or_vector_selector<-1, T> {
 788 | 		typedef std::vector<T> container_t;
 789 | 	};
 790 | 	
 791 | 	/** @} */
 792 | 
 793 | 	/** kd-tree base-class
 794 | 	 *
 795 | 	 * Contains the member functions common to the classes KDTreeSingleIndexAdaptor and KDTreeSingleIndexDynamicAdaptor_.
 796 | 	 *
 797 | 	 * \tparam Derived The name of the class which inherits this class.
 798 | 	 * \tparam DatasetAdaptor The user-provided adaptor (see comments above).
 799 | 	 * \tparam Distance The distance metric to use, these are all classes derived from nanoflann::Metric
 800 | 	 * \tparam DIM Dimensionality of data points (e.g. 3 for 3D points)
 801 | 	 * \tparam IndexType Will be typically size_t or int
 802 | 	 */
 803 | 
 804 | 	template<class Derived, typename Distance, class DatasetAdaptor, int DIM = -1, typename IndexType = size_t>
 805 | 	class KDTreeBaseClass
 806 | 	{
 807 | 
 808 | 	public:
 809 | 		/** Frees the previously-built index. Automatically called within buildIndex(). */
 810 | 		void freeIndex(Derived &obj)
 811 | 		{
 812 | 			obj.pool.free_all();
 813 | 			obj.root_node = NULL;
 814 | 			obj.m_size_at_index_build = 0;
 815 | 		}
 816 | 
 817 | 		typedef typename Distance::ElementType  ElementType;
 818 | 		typedef typename Distance::DistanceType DistanceType;
 819 | 
 820 | 		/*--------------------- Internal Data Structures --------------------------*/
 821 | 		struct Node
 822 | 		{
 823 | 			/** Union used because a node can be either a LEAF node or a non-leaf node, so both data fields are never used simultaneously */
 824 | 			union {
 825 | 				struct leaf
 826 |                                 {
 827 | 					IndexType    left, right;  //!< Indices of points in leaf node
 828 | 				} lr;
 829 | 				struct nonleaf
 830 |                                 {
 831 | 					int          divfeat; //!< Dimension used for subdivision.
 832 | 					DistanceType divlow, divhigh; //!< The values used for subdivision.
 833 | 				} sub;
 834 | 			} node_type;
 835 | 			Node *child1, *child2;  //!< Child nodes (both=NULL mean its a leaf node)
 836 | 		};
 837 | 		
 838 | 		typedef Node* NodePtr;
 839 | 
 840 | 		struct Interval
 841 | 		{
 842 | 			ElementType low, high;
 843 | 		};
 844 | 
 845 | 		/**
 846 | 		 *  Array of indices to vectors in the dataset.
 847 | 		 */
 848 | 		std::vector<IndexType> vind;
 849 | 
 850 | 		NodePtr root_node;
 851 | 
 852 | 		size_t m_leaf_max_size;
 853 | 
 854 | 		size_t m_size; //!< Number of current points in the dataset
 855 | 		size_t m_size_at_index_build; //!< Number of points in the dataset when the index was built
 856 | 		int dim;  //!< Dimensionality of each data point
 857 | 
 858 | 		/** Define "BoundingBox" as a fixed-size or variable-size container depending on "DIM" */
 859 | 		typedef typename array_or_vector_selector<DIM, Interval>::container_t BoundingBox;
 860 | 
 861 | 		/** Define "distance_vector_t" as a fixed-size or variable-size container depending on "DIM" */
 862 | 		typedef typename array_or_vector_selector<DIM, DistanceType>::container_t distance_vector_t;
 863 | 
 864 | 		/** The KD-tree used to find neighbours */
 865 | 		
 866 | 		BoundingBox root_bbox;
 867 | 
 868 | 		/**
 869 | 		 * Pooled memory allocator.
 870 | 		 *
 871 | 		 * Using a pooled memory allocator is more efficient
 872 | 		 * than allocating memory directly when there is a large
 873 | 		 * number small of memory allocations.
 874 | 		 */
 875 | 		PooledAllocator pool;
 876 | 
 877 | 		/** Returns number of points in dataset  */
 878 | 		size_t size(const Derived &obj) const { return obj.m_size; }
 879 | 
 880 | 		/** Returns the length of each point in the dataset */
 881 | 		size_t veclen(const Derived &obj) {
 882 | 			return static_cast<size_t>(DIM>0 ? DIM : obj.dim);
 883 | 		}
 884 | 
 885 | 		/// Helper accessor to the dataset points:
 886 | 		inline ElementType dataset_get(const Derived &obj, size_t idx, int component) const{
 887 | 			return obj.dataset.kdtree_get_pt(idx, component);
 888 | 		}
 889 | 
 890 | 		/**
 891 | 		 * Computes the inde memory usage
 892 | 		 * Returns: memory used by the index
 893 | 		 */
 894 | 		size_t usedMemory(Derived &obj)
 895 | 		{
 896 | 			return obj.pool.usedMemory + obj.pool.wastedMemory + obj.dataset.kdtree_get_point_count() * sizeof(IndexType);  // pool memory and vind array memory
 897 | 		}
 898 | 
 899 | 		void computeMinMax(const Derived &obj, IndexType* ind, IndexType count, int element, ElementType& min_elem, ElementType& max_elem)
 900 | 		{
 901 | 			min_elem = dataset_get(obj, ind[0],element);
 902 | 			max_elem = dataset_get(obj, ind[0],element);
 903 | 			for (IndexType i = 1; i < count; ++i) {
 904 | 				ElementType val = dataset_get(obj, ind[i], element);
 905 | 				if (val < min_elem) min_elem = val;
 906 | 				if (val > max_elem) max_elem = val;
 907 | 			}
 908 | 		}
 909 | 
 910 | 		/**
 911 | 		 * Create a tree node that subdivides the list of vecs from vind[first]
 912 | 		 * to vind[last].  The routine is called recursively on each sublist.
 913 | 		 *
 914 | 		 * @param left index of the first vector
 915 | 		 * @param right index of the last vector
 916 | 		 */
 917 | 		NodePtr divideTree(Derived &obj, const IndexType left, const IndexType right, BoundingBox& bbox)
 918 | 		{
 919 | 			NodePtr node = obj.pool.template allocate<Node>(); // allocate memory
 920 | 
 921 | 			/* If too few exemplars remain, then make this a leaf node. */
 922 | 			if ( (right - left) <= static_cast<IndexType>(obj.m_leaf_max_size) ) {
 923 | 				node->child1 = node->child2 = NULL;    /* Mark as leaf node. */
 924 | 				node->node_type.lr.left = left;
 925 | 				node->node_type.lr.right = right;
 926 | 
 927 | 				// compute bounding-box of leaf points
 928 | 				for (int i = 0; i < (DIM > 0 ? DIM : obj.dim); ++i) {
 929 | 					bbox[i].low = dataset_get(obj, obj.vind[left], i);
 930 | 					bbox[i].high = dataset_get(obj, obj.vind[left], i);
 931 | 				}
 932 | 				for (IndexType k = left + 1; k < right; ++k) {
 933 | 					for (int i = 0; i < (DIM > 0 ? DIM : obj.dim); ++i) {
 934 | 						if (bbox[i].low > dataset_get(obj, obj.vind[k], i)) bbox[i].low = dataset_get(obj, obj.vind[k], i);
 935 | 						if (bbox[i].high < dataset_get(obj, obj.vind[k], i)) bbox[i].high = dataset_get(obj, obj.vind[k], i);
 936 | 					}
 937 | 				}
 938 | 			}
 939 | 			else {
 940 | 				IndexType idx;
 941 | 				int cutfeat;
 942 | 				DistanceType cutval;
 943 | 				middleSplit_(obj, &obj.vind[0] + left, right - left, idx, cutfeat, cutval, bbox);
 944 | 
 945 | 				node->node_type.sub.divfeat = cutfeat;
 946 | 
 947 | 				BoundingBox left_bbox(bbox);
 948 | 				left_bbox[cutfeat].high = cutval;
 949 | 				node->child1 = divideTree(obj, left, left + idx, left_bbox);
 950 | 
 951 | 				BoundingBox right_bbox(bbox);
 952 | 				right_bbox[cutfeat].low = cutval;
 953 | 				node->child2 = divideTree(obj, left + idx, right, right_bbox);
 954 | 
 955 | 				node->node_type.sub.divlow = left_bbox[cutfeat].high;
 956 | 				node->node_type.sub.divhigh = right_bbox[cutfeat].low;
 957 | 
 958 | 				for (int i = 0; i < (DIM > 0 ? DIM : obj.dim); ++i) {
 959 | 					bbox[i].low = std::min(left_bbox[i].low, right_bbox[i].low);
 960 | 					bbox[i].high = std::max(left_bbox[i].high, right_bbox[i].high);
 961 | 				}
 962 | 			}
 963 | 
 964 | 			return node;
 965 | 		}
 966 | 
 967 | 		void middleSplit_(Derived &obj, IndexType* ind, IndexType count, IndexType& index, int& cutfeat, DistanceType& cutval, const BoundingBox& bbox)
 968 | 		{
 969 | 			const DistanceType EPS = static_cast<DistanceType>(0.00001);
 970 | 			ElementType max_span = bbox[0].high-bbox[0].low;
 971 | 			for (int i = 1; i < (DIM > 0 ? DIM : obj.dim); ++i) {
 972 | 				ElementType span = bbox[i].high - bbox[i].low;
 973 | 				if (span > max_span) {
 974 | 					max_span = span;
 975 | 				}
 976 | 			}
 977 | 			ElementType max_spread = -1;
 978 | 			cutfeat = 0;
 979 | 			for (int i = 0; i < (DIM > 0 ? DIM : obj.dim); ++i) {
 980 | 				ElementType span = bbox[i].high-bbox[i].low;
 981 | 				if (span > (1 - EPS) * max_span) {
 982 | 					ElementType min_elem, max_elem;
 983 | 					computeMinMax(obj, ind, count, i, min_elem, max_elem);
 984 | 					ElementType spread = max_elem - min_elem;;
 985 | 					if (spread > max_spread) {
 986 | 						cutfeat = i;
 987 | 						max_spread = spread;
 988 | 					}
 989 | 				}
 990 | 			}
 991 | 			// split in the middle
 992 | 			DistanceType split_val = (bbox[cutfeat].low + bbox[cutfeat].high) / 2;
 993 | 			ElementType min_elem, max_elem;
 994 | 			computeMinMax(obj, ind, count, cutfeat, min_elem, max_elem);
 995 | 
 996 | 			if (split_val < min_elem) cutval = min_elem;
 997 | 			else if (split_val > max_elem) cutval = max_elem;
 998 | 			else cutval = split_val;
 999 | 
1000 | 			IndexType lim1, lim2;
1001 | 			planeSplit(obj, ind, count, cutfeat, cutval, lim1, lim2);
1002 | 
1003 | 			if (lim1 > count / 2) index = lim1;
1004 | 			else if (lim2 < count / 2) index = lim2;
1005 | 			else index = count/2;
1006 | 		}
1007 | 
1008 | 		/**
1009 | 		 *  Subdivide the list of points by a plane perpendicular on axe corresponding
1010 | 		 *  to the 'cutfeat' dimension at 'cutval' position.
1011 | 		 *
1012 | 		 *  On return:
1013 | 		 *  dataset[ind[0..lim1-1]][cutfeat]<cutval
1014 | 		 *  dataset[ind[lim1..lim2-1]][cutfeat]==cutval
1015 | 		 *  dataset[ind[lim2..count]][cutfeat]>cutval
1016 | 		 */
1017 | 		void planeSplit(Derived &obj, IndexType* ind, const IndexType count, int cutfeat, DistanceType &cutval, IndexType& lim1, IndexType& lim2)
1018 | 		{
1019 | 			/* Move vector indices for left subtree to front of list. */
1020 | 			IndexType left = 0;
1021 | 			IndexType right = count-1;
1022 | 			for (;; ) {
1023 | 				while (left <= right && dataset_get(obj, ind[left], cutfeat) < cutval) ++left;
1024 | 				while (right && left <= right && dataset_get(obj, ind[right], cutfeat) >= cutval) --right;
1025 | 				if (left > right || !right) break;  // "!right" was added to support unsigned Index types
1026 | 				std::swap(ind[left], ind[right]);
1027 | 				++left;
1028 | 				--right;
1029 | 			}
1030 | 			/* If either list is empty, it means that all remaining features
1031 | 			 * are identical. Split in the middle to maintain a balanced tree.
1032 | 			 */
1033 | 			lim1 = left;
1034 | 			right = count-1;
1035 | 			for (;; ) {
1036 | 				while (left <= right && dataset_get(obj, ind[left], cutfeat) <= cutval) ++left;
1037 | 				while (right && left <= right && dataset_get(obj, ind[right], cutfeat) > cutval) --right;
1038 | 				if (left > right || !right) break;  // "!right" was added to support unsigned Index types
1039 | 				std::swap(ind[left], ind[right]);
1040 | 				++left;
1041 | 				--right;
1042 | 			}
1043 | 			lim2 = left;
1044 | 		}
1045 | 
1046 | 		DistanceType computeInitialDistances(const Derived &obj, const ElementType* vec, distance_vector_t& dists) const
1047 | 		{
1048 | 			assert(vec);
1049 | 			DistanceType distsq = DistanceType();
1050 | 
1051 | 			for (int i = 0; i < (DIM>0 ? DIM : obj.dim); ++i) {
1052 | 				if (vec[i] < obj.root_bbox[i].low) {
1053 | 					dists[i] = obj.distance.accum_dist(vec[i], obj.root_bbox[i].low, i);
1054 | 					distsq += dists[i];
1055 | 				}
1056 | 				if (vec[i] > obj.root_bbox[i].high) {
1057 | 					dists[i] = obj.distance.accum_dist(vec[i], obj.root_bbox[i].high, i);
1058 | 					distsq += dists[i];
1059 | 				}
1060 | 			}
1061 | 			return distsq;
1062 | 		}
1063 | 
1064 | 		void save_tree(Derived &obj, FILE* stream, NodePtr tree)
1065 | 		{
1066 | 			save_value(stream, *tree);
1067 | 			if (tree->child1 != NULL) {
1068 | 				save_tree(obj, stream, tree->child1);
1069 | 			}
1070 | 			if (tree->child2 != NULL) {
1071 | 				save_tree(obj, stream, tree->child2);
1072 | 			}
1073 | 		}
1074 | 
1075 | 
1076 | 		void load_tree(Derived &obj, FILE* stream, NodePtr& tree)
1077 | 		{
1078 | 			tree = obj.pool.template allocate<Node>();
1079 | 			load_value(stream, *tree);
1080 | 			if (tree->child1 != NULL) {
1081 | 				load_tree(obj, stream, tree->child1);
1082 | 			}
1083 | 			if (tree->child2 != NULL) {
1084 | 				load_tree(obj, stream, tree->child2);
1085 | 			}
1086 | 		}
1087 | 
1088 | 		/**  Stores the index in a binary file.
1089 | 		  *   IMPORTANT NOTE: The set of data points is NOT stored in the file, so when loading the index object it must be constructed associated to the same source of data points used while building it.
1090 | 		  * See the example: examples/saveload_example.cpp
1091 | 		  * \sa loadIndex  */
1092 | 		void saveIndex_(Derived &obj, FILE* stream)
1093 | 		{
1094 | 			save_value(stream, obj.m_size);
1095 | 			save_value(stream, obj.dim);
1096 | 			save_value(stream, obj.root_bbox);
1097 | 			save_value(stream, obj.m_leaf_max_size);
1098 | 			save_value(stream, obj.vind);
1099 | 			save_tree(obj, stream, obj.root_node);
1100 | 		}
1101 | 
1102 | 		/**  Loads a previous index from a binary file.
1103 | 		  *   IMPORTANT NOTE: The set of data points is NOT stored in the file, so the index object must be constructed associated to the same source of data points used while building the index.
1104 | 		  * See the example: examples/saveload_example.cpp
1105 | 		  * \sa loadIndex  */
1106 | 		void loadIndex_(Derived &obj, FILE* stream)
1107 | 		{
1108 | 			load_value(stream, obj.m_size);
1109 | 			load_value(stream, obj.dim);
1110 | 			load_value(stream, obj.root_bbox);
1111 | 			load_value(stream, obj.m_leaf_max_size);
1112 | 			load_value(stream, obj.vind);
1113 | 			load_tree(obj, stream, obj.root_node);
1114 | 		}
1115 | 
1116 | 	};
1117 | 
1118 | 
1119 | 	/** @addtogroup kdtrees_grp KD-tree classes and adaptors
1120 | 	  * @{ */
1121 | 
1122 | 	/** kd-tree static index
1123 | 	 *
1124 | 	 * Contains the k-d trees and other information for indexing a set of points
1125 | 	 * for nearest-neighbor matching.
1126 | 	 *
1127 | 	 *  The class "DatasetAdaptor" must provide the following interface (can be non-virtual, inlined methods):
1128 | 	 *
1129 | 	 *  \code
1130 | 	 *   // Must return the number of data poins
1131 | 	 *   inline size_t kdtree_get_point_count() const { ... }
1132 | 	 *
1133 | 	 *
1134 | 	 *   // Must return the dim'th component of the idx'th point in the class:
1135 | 	 *   inline T kdtree_get_pt(const size_t idx, int dim) const { ... }
1136 | 	 *
1137 | 	 *   // Optional bounding-box computation: return false to default to a standard bbox computation loop.
1138 | 	 *   //   Return true if the BBOX was already computed by the class and returned in "bb" so it can be avoided to redo it again.
1139 | 	 *   //   Look at bb.size() to find out the expected dimensionality (e.g. 2 or 3 for point clouds)
1140 | 	 *   template <class BBOX>
1141 | 	 *   bool kdtree_get_bbox(BBOX &bb) const
1142 | 	 *   {
1143 | 	 *      bb[0].low = ...; bb[0].high = ...;  // 0th dimension limits
1144 | 	 *      bb[1].low = ...; bb[1].high = ...;  // 1st dimension limits
1145 | 	 *      ...
1146 | 	 *      return true;
1147 | 	 *   }
1148 | 	 *
1149 | 	 *  \endcode
1150 | 	 *
1151 | 	 * \tparam DatasetAdaptor The user-provided adaptor (see comments above).
1152 | 	 * \tparam Distance The distance metric to use: nanoflann::metric_L1, nanoflann::metric_L2, nanoflann::metric_L2_Simple, etc.
1153 | 	 * \tparam DIM Dimensionality of data points (e.g. 3 for 3D points)
1154 | 	 * \tparam IndexType Will be typically size_t or int
1155 | 	 */
1156 | 	template <typename Distance, class DatasetAdaptor, int DIM = -1, typename IndexType = size_t>
1157 | 	class KDTreeSingleIndexAdaptor : public KDTreeBaseClass<KDTreeSingleIndexAdaptor<Distance, DatasetAdaptor, DIM, IndexType>, Distance, DatasetAdaptor, DIM, IndexType>
1158 | 	{
1159 | 	private:
1160 | 		/** Hidden copy constructor, to disallow copying indices (Not implemented) */
1161 | 		KDTreeSingleIndexAdaptor(const KDTreeSingleIndexAdaptor<Distance, DatasetAdaptor, DIM, IndexType>&);
1162 | 	public:
1163 | 		
1164 | 		/**
1165 | 		 * The dataset used by this index
1166 | 		 */
1167 | 		const DatasetAdaptor &dataset; //!< The source of our data
1168 | 
1169 | 		const KDTreeSingleIndexAdaptorParams index_params;
1170 | 
1171 | 		Distance distance;
1172 | 
1173 | 		typedef typename nanoflann::KDTreeBaseClass<nanoflann::KDTreeSingleIndexAdaptor<Distance, DatasetAdaptor, DIM, IndexType>, Distance, DatasetAdaptor, DIM, IndexType> BaseClassRef;
1174 | 
1175 | 		typedef typename BaseClassRef::ElementType ElementType;
1176 | 		typedef typename BaseClassRef::DistanceType DistanceType;
1177 | 
1178 | 		typedef typename BaseClassRef::Node Node;
1179 | 		typedef Node* NodePtr;
1180 | 
1181 | 		typedef typename BaseClassRef::Interval Interval;
1182 | 		/** Define "BoundingBox" as a fixed-size or variable-size container depending on "DIM" */
1183 | 		typedef typename BaseClassRef::BoundingBox BoundingBox;
1184 | 
1185 | 		/** Define "distance_vector_t" as a fixed-size or variable-size container depending on "DIM" */
1186 | 		typedef typename BaseClassRef::distance_vector_t distance_vector_t;
1187 | 
1188 | 		/**
1189 | 		 * KDTree constructor
1190 | 		 *
1191 | 		 * Refer to docs in README.md or online in https://github.com/jlblancoc/nanoflann
1192 | 		 *
1193 | 		 * The KD-Tree point dimension (the length of each point in the datase, e.g. 3 for 3D points)
1194 | 		 * is determined by means of:
1195 | 		 *  - The \a DIM template parameter if >0 (highest priority)
1196 | 		 *  - Otherwise, the \a dimensionality parameter of this constructor.
1197 | 		 *
1198 | 		 * @param inputData Dataset with the input features
1199 | 		 * @param params Basically, the maximum leaf node size
1200 | 		 */
1201 | 		KDTreeSingleIndexAdaptor(const int dimensionality, const DatasetAdaptor& inputData, const KDTreeSingleIndexAdaptorParams& params = KDTreeSingleIndexAdaptorParams() ) :
1202 | 			dataset(inputData), index_params(params), distance(inputData)
1203 | 		{
1204 | 			BaseClassRef::root_node = NULL;
1205 | 			BaseClassRef::m_size = dataset.kdtree_get_point_count();
1206 | 			BaseClassRef::m_size_at_index_build = BaseClassRef::m_size;
1207 | 			BaseClassRef::dim = dimensionality;
1208 | 			if (DIM>0) BaseClassRef::dim = DIM;
1209 | 			BaseClassRef::m_leaf_max_size = params.leaf_max_size;
1210 | 
1211 | 			// Create a permutable array of indices to the input vectors.
1212 | 			init_vind();
1213 | 		}
1214 | 
1215 | 		/**
1216 | 		 * Builds the index
1217 | 		 */
1218 | 		void buildIndex()
1219 | 		{
1220 | 			BaseClassRef::m_size = dataset.kdtree_get_point_count();
1221 | 			BaseClassRef::m_size_at_index_build = BaseClassRef::m_size;
1222 | 			init_vind();
1223 | 			this->freeIndex(*this);
1224 | 			BaseClassRef::m_size_at_index_build = BaseClassRef::m_size;
1225 | 			if(BaseClassRef::m_size == 0) return;
1226 | 			computeBoundingBox(BaseClassRef::root_bbox);
1227 | 			BaseClassRef::root_node = this->divideTree(*this, 0, BaseClassRef::m_size, BaseClassRef::root_bbox );   // construct the tree
1228 | 		}
1229 | 
1230 | 		/** \name Query methods
1231 | 		  * @{ */
1232 | 
1233 | 		/**
1234 | 		 * Find set of nearest neighbors to vec[0:dim-1]. Their indices are stored inside
1235 | 		 * the result object.
1236 | 		 *
1237 | 		 * Params:
1238 | 		 *     result = the result object in which the indices of the nearest-neighbors are stored
1239 | 		 *     vec = the vector for which to search the nearest neighbors
1240 | 		 *
1241 | 		 * \tparam RESULTSET Should be any ResultSet<DistanceType>
1242 |          * \return  True if the requested neighbors could be found.
1243 | 		 * \sa knnSearch, radiusSearch
1244 | 		 */
1245 | 		template <typename RESULTSET>
1246 | 		bool findNeighbors(RESULTSET& result, const ElementType* vec, const SearchParams& searchParams) const
1247 | 		{
1248 | 			assert(vec);
1249 |             if (this->size(*this) == 0)
1250 |                 return false;
1251 | 			if (!BaseClassRef::root_node)
1252 |                 throw std::runtime_error("[nanoflann] findNeighbors() called before building the index.");
1253 | 			float epsError = 1 + searchParams.eps;
1254 | 
1255 | 			distance_vector_t dists; // fixed or variable-sized container (depending on DIM)
1256 | 			dists.assign((DIM > 0 ? DIM : BaseClassRef::dim), 0); // Fill it with zeros.
1257 | 			DistanceType distsq = this->computeInitialDistances(*this, vec, dists);
1258 | 			searchLevel(result, vec, BaseClassRef::root_node, distsq, dists, epsError);  // "count_leaf" parameter removed since was neither used nor returned to the user.
1259 |             return result.full();
1260 | 		}
1261 | 
1262 | 		/**
1263 | 		 * Find the "num_closest" nearest neighbors to the \a query_point[0:dim-1]. Their indices are stored inside
1264 | 		 * the result object.
1265 | 		 *  \sa radiusSearch, findNeighbors
1266 | 		 * \note nChecks_IGNORED is ignored but kept for compatibility with the original FLANN interface.
1267 | 		 * \return Number `N` of valid points in the result set. Only the first `N` entries in `out_indices` and `out_distances_sq` will be valid.
1268 | 		 *         Return may be less than `num_closest` only if the number of elements in the tree is less than `num_closest`.
1269 | 		 */
1270 | 		size_t knnSearch(const ElementType *query_point, const size_t num_closest, IndexType *out_indices, DistanceType *out_distances_sq, const int /* nChecks_IGNORED */ = 10) const
1271 | 		{
1272 | 			nanoflann::KNNResultSet<DistanceType,IndexType> resultSet(num_closest);
1273 | 			resultSet.init(out_indices, out_distances_sq);
1274 | 			this->findNeighbors(resultSet, query_point, nanoflann::SearchParams());
1275 | 			return resultSet.size();
1276 | 		}
1277 | 
1278 | 		/**
1279 | 		 * Find all the neighbors to \a query_point[0:dim-1] within a maximum radius.
1280 | 		 *  The output is given as a vector of pairs, of which the first element is a point index and the second the corresponding distance.
1281 | 		 *  Previous contents of \a IndicesDists are cleared.
1282 | 		 *
1283 | 		 *  If searchParams.sorted==true, the output list is sorted by ascending distances.
1284 | 		 *
1285 | 		 *  For a better performance, it is advisable to do a .reserve() on the vector if you have any wild guess about the number of expected matches.
1286 | 		 *
1287 | 		 *  \sa knnSearch, findNeighbors, radiusSearchCustomCallback
1288 | 		 * \return The number of points within the given radius (i.e. indices.size() or dists.size() )
1289 | 		 */
1290 | 		size_t radiusSearch(const ElementType *query_point, const DistanceType &radius, std::vector<std::pair<IndexType, DistanceType> >& IndicesDists, const SearchParams& searchParams) const
1291 | 		{
1292 | 			RadiusResultSet<DistanceType, IndexType> resultSet(radius, IndicesDists);
1293 | 			const size_t nFound = radiusSearchCustomCallback(query_point, resultSet, searchParams);
1294 | 			if (searchParams.sorted)
1295 | 				std::sort(IndicesDists.begin(), IndicesDists.end(), IndexDist_Sorter() );
1296 | 			return nFound;
1297 | 		}
1298 | 
1299 | 		/**
1300 | 		 * Just like radiusSearch() but with a custom callback class for each point found in the radius of the query.
1301 | 		 * See the source of RadiusResultSet<> as a start point for your own classes.
1302 | 		 * \sa radiusSearch
1303 | 		 */
1304 | 		template <class SEARCH_CALLBACK>
1305 | 		size_t radiusSearchCustomCallback(const ElementType *query_point, SEARCH_CALLBACK &resultSet, const SearchParams& searchParams = SearchParams() ) const
1306 | 		{
1307 | 			this->findNeighbors(resultSet, query_point, searchParams);
1308 | 			return resultSet.size();
1309 | 		}
1310 | 
1311 | 		/** @} */
1312 | 
1313 | 	public:
1314 | 		/** Make sure the auxiliary list \a vind has the same size than the current dataset, and re-generate if size has changed. */
1315 | 		void init_vind()
1316 | 		{
1317 | 			// Create a permutable array of indices to the input vectors.
1318 | 			BaseClassRef::m_size = dataset.kdtree_get_point_count();
1319 | 			if (BaseClassRef::vind.size() != BaseClassRef::m_size) BaseClassRef::vind.resize(BaseClassRef::m_size);
1320 | 			for (size_t i = 0; i < BaseClassRef::m_size; i++) BaseClassRef::vind[i] = i;
1321 | 		}
1322 | 
1323 | 		void computeBoundingBox(BoundingBox& bbox)
1324 | 		{
1325 | 			bbox.resize((DIM > 0 ? DIM : BaseClassRef::dim));
1326 | 			if (dataset.kdtree_get_bbox(bbox))
1327 | 			{
1328 | 				// Done! It was implemented in derived class
1329 | 			}
1330 | 			else
1331 | 			{
1332 | 				const size_t N = dataset.kdtree_get_point_count();
1333 | 				if (!N) throw std::runtime_error("[nanoflann] computeBoundingBox() called but no data points found.");
1334 | 				for (int i = 0; i < (DIM > 0 ? DIM : BaseClassRef::dim); ++i) {
1335 | 					bbox[i].low =
1336 | 					bbox[i].high = this->dataset_get(*this, 0, i);
1337 | 				}
1338 | 				for (size_t k = 1; k < N; ++k) {
1339 | 					for (int i = 0; i < (DIM > 0 ? DIM : BaseClassRef::dim); ++i) {
1340 | 						if (this->dataset_get(*this, k, i) < bbox[i].low) bbox[i].low = this->dataset_get(*this, k, i);
1341 | 						if (this->dataset_get(*this, k, i) > bbox[i].high) bbox[i].high = this->dataset_get(*this, k, i);
1342 | 					}
1343 | 				}
1344 | 			}
1345 | 		}
1346 | 
1347 | 		/**
1348 | 		 * Performs an exact search in the tree starting from a node.
1349 | 		 * \tparam RESULTSET Should be any ResultSet<DistanceType>
1350 |                  * \return true if the search should be continued, false if the results are sufficient
1351 | 		 */
1352 | 		template <class RESULTSET>
1353 |                 bool searchLevel(RESULTSET& result_set, const ElementType* vec, const NodePtr node, DistanceType mindistsq,
1354 | 						 distance_vector_t& dists, const float epsError) const
1355 | 		{
1356 | 			/* If this is a leaf node, then do check and return. */
1357 | 			if ((node->child1 == NULL) && (node->child2 == NULL)) {
1358 | 				//count_leaf += (node->lr.right-node->lr.left);  // Removed since was neither used nor returned to the user.
1359 | 				DistanceType worst_dist = result_set.worstDist();
1360 | 				for (IndexType i = node->node_type.lr.left; i<node->node_type.lr.right; ++i) {
1361 | 					const IndexType index = BaseClassRef::vind[i];// reorder... : i;
1362 | 					DistanceType dist = distance.evalMetric(vec, index, (DIM > 0 ? DIM : BaseClassRef::dim));
1363 | 					if (dist < worst_dist) {
1364 |                                                 if(!result_set.addPoint(dist, BaseClassRef::vind[i])) {
1365 |                                                     // the resultset doesn't want to receive any more points, we're done searching!
1366 |                                                     return false;
1367 |                                                 }
1368 | 					}
1369 | 				}
1370 |                                 return true;
1371 | 			}
1372 | 
1373 | 			/* Which child branch should be taken first? */
1374 | 			int idx = node->node_type.sub.divfeat;
1375 | 			ElementType val = vec[idx];
1376 | 			DistanceType diff1 = val - node->node_type.sub.divlow;
1377 | 			DistanceType diff2 = val - node->node_type.sub.divhigh;
1378 | 
1379 | 			NodePtr bestChild;
1380 | 			NodePtr otherChild;
1381 | 			DistanceType cut_dist;
1382 | 			if ((diff1 + diff2) < 0) {
1383 | 				bestChild = node->child1;
1384 | 				otherChild = node->child2;
1385 | 				cut_dist = distance.accum_dist(val, node->node_type.sub.divhigh, idx);
1386 | 			}
1387 | 			else {
1388 | 				bestChild = node->child2;
1389 | 				otherChild = node->child1;
1390 | 				cut_dist = distance.accum_dist( val, node->node_type.sub.divlow, idx);
1391 | 			}
1392 | 
1393 | 			/* Call recursively to search next level down. */
1394 |                         if(!searchLevel(result_set, vec, bestChild, mindistsq, dists, epsError)) {
1395 |                             // the resultset doesn't want to receive any more points, we're done searching!
1396 |                             return false;
1397 |                         }
1398 | 
1399 | 			DistanceType dst = dists[idx];
1400 | 			mindistsq = mindistsq + cut_dist - dst;
1401 | 			dists[idx] = cut_dist;
1402 | 			if (mindistsq*epsError <= result_set.worstDist()) {
1403 |                             if(!searchLevel(result_set, vec, otherChild, mindistsq, dists, epsError)) {
1404 |                                 // the resultset doesn't want to receive any more points, we're done searching!
1405 |                                 return false;
1406 |                             }
1407 | 			}
1408 | 			dists[idx] = dst;
1409 |                         return true;
1410 | 		}
1411 | 
1412 | 	public:
1413 | 		/**  Stores the index in a binary file.
1414 | 		  *   IMPORTANT NOTE: The set of data points is NOT stored in the file, so when loading the index object it must be constructed associated to the same source of data points used while building it.
1415 | 		  * See the example: examples/saveload_example.cpp
1416 | 		  * \sa loadIndex  */
1417 | 		void saveIndex(FILE* stream)
1418 | 		{
1419 | 			this->saveIndex_(*this, stream);
1420 | 		}
1421 | 
1422 | 		/**  Loads a previous index from a binary file.
1423 | 		  *   IMPORTANT NOTE: The set of data points is NOT stored in the file, so the index object must be constructed associated to the same source of data points used while building the index.
1424 | 		  * See the example: examples/saveload_example.cpp
1425 | 		  * \sa loadIndex  */
1426 | 		void loadIndex(FILE* stream)
1427 | 		{
1428 | 			this->loadIndex_(*this, stream);
1429 | 		}
1430 | 
1431 | 	};   // class KDTree
1432 | 
1433 | 
1434 | 	/** kd-tree dynamic index
1435 | 	 *
1436 | 	 * Contains the k-d trees and other information for indexing a set of points
1437 | 	 * for nearest-neighbor matching.
1438 | 	 *
1439 | 	 *  The class "DatasetAdaptor" must provide the following interface (can be non-virtual, inlined methods):
1440 | 	 *
1441 | 	 *  \code
1442 | 	 *   // Must return the number of data poins
1443 | 	 *   inline size_t kdtree_get_point_count() const { ... }
1444 | 	 *
1445 | 	 *   // Must return the dim'th component of the idx'th point in the class:
1446 | 	 *   inline T kdtree_get_pt(const size_t idx, int dim) const { ... }
1447 | 	 *
1448 | 	 *   // Optional bounding-box computation: return false to default to a standard bbox computation loop.
1449 | 	 *   //   Return true if the BBOX was already computed by the class and returned in "bb" so it can be avoided to redo it again.
1450 | 	 *   //   Look at bb.size() to find out the expected dimensionality (e.g. 2 or 3 for point clouds)
1451 | 	 *   template <class BBOX>
1452 | 	 *   bool kdtree_get_bbox(BBOX &bb) const
1453 | 	 *   {
1454 | 	 *      bb[0].low = ...; bb[0].high = ...;  // 0th dimension limits
1455 | 	 *      bb[1].low = ...; bb[1].high = ...;  // 1st dimension limits
1456 | 	 *      ...
1457 | 	 *      return true;
1458 | 	 *   }
1459 | 	 *
1460 | 	 *  \endcode
1461 | 	 *
1462 | 	 * \tparam DatasetAdaptor The user-provided adaptor (see comments above).
1463 | 	 * \tparam Distance The distance metric to use: nanoflann::metric_L1, nanoflann::metric_L2, nanoflann::metric_L2_Simple, etc.
1464 | 	 * \tparam DIM Dimensionality of data points (e.g. 3 for 3D points)
1465 | 	 * \tparam IndexType Will be typically size_t or int
1466 | 	 */	 
1467 | 	template <typename Distance, class DatasetAdaptor, int DIM = -1, typename IndexType = size_t>
1468 | 	class KDTreeSingleIndexDynamicAdaptor_ : public KDTreeBaseClass<KDTreeSingleIndexDynamicAdaptor_<Distance, DatasetAdaptor, DIM, IndexType>, Distance, DatasetAdaptor, DIM, IndexType>
1469 | 	{
1470 | 	public:
1471 | 
1472 | 		/**
1473 | 		 * The dataset used by this index
1474 | 		 */
1475 | 		DatasetAdaptor &dataset; //!< The source of our data
1476 | 
1477 | 		KDTreeSingleIndexAdaptorParams index_params;
1478 | 
1479 | 		std::vector<int> &treeIndex;
1480 | 
1481 | 		Distance distance;
1482 | 
1483 | 		typedef typename nanoflann::KDTreeBaseClass<nanoflann::KDTreeSingleIndexDynamicAdaptor_<Distance, DatasetAdaptor, DIM, IndexType>, Distance, DatasetAdaptor, DIM, IndexType> BaseClassRef;
1484 | 
1485 | 		typedef typename BaseClassRef::ElementType  ElementType;
1486 | 		typedef typename BaseClassRef::DistanceType  DistanceType;
1487 | 
1488 | 		typedef typename BaseClassRef::Node Node;
1489 | 		typedef Node* NodePtr;
1490 | 
1491 | 		typedef typename BaseClassRef::Interval Interval;
1492 | 		/** Define "BoundingBox" as a fixed-size or variable-size container depending on "DIM" */
1493 | 		typedef typename BaseClassRef::BoundingBox BoundingBox;
1494 | 
1495 | 		/** Define "distance_vector_t" as a fixed-size or variable-size container depending on "DIM" */
1496 | 		typedef typename BaseClassRef::distance_vector_t distance_vector_t;
1497 | 
1498 | 		/**
1499 | 		 * KDTree constructor
1500 | 		 *
1501 | 		 * Refer to docs in README.md or online in https://github.com/jlblancoc/nanoflann
1502 | 		 *
1503 | 		 * The KD-Tree point dimension (the length of each point in the datase, e.g. 3 for 3D points)
1504 | 		 * is determined by means of:
1505 | 		 *  - The \a DIM template parameter if >0 (highest priority)
1506 | 		 *  - Otherwise, the \a dimensionality parameter of this constructor.
1507 | 		 *
1508 | 		 * @param inputData Dataset with the input features
1509 | 		 * @param params Basically, the maximum leaf node size
1510 | 		 */
1511 | 		KDTreeSingleIndexDynamicAdaptor_(const int dimensionality, DatasetAdaptor& inputData, std::vector<int>& treeIndex_, const KDTreeSingleIndexAdaptorParams& params = KDTreeSingleIndexAdaptorParams()) :
1512 | 			dataset(inputData), index_params(params), treeIndex(treeIndex_), distance(inputData)
1513 | 		{
1514 | 			BaseClassRef::root_node = NULL;
1515 | 			BaseClassRef::m_size = 0;
1516 | 			BaseClassRef::m_size_at_index_build = 0;
1517 | 			BaseClassRef::dim = dimensionality;
1518 | 			if (DIM>0) BaseClassRef::dim = DIM;
1519 | 			BaseClassRef::m_leaf_max_size = params.leaf_max_size;
1520 | 		}
1521 | 
1522 | 
1523 | 		/** Assignment operator definiton */
1524 | 		KDTreeSingleIndexDynamicAdaptor_ operator=( const KDTreeSingleIndexDynamicAdaptor_& rhs ) {
1525 | 		      KDTreeSingleIndexDynamicAdaptor_ tmp( rhs );
1526 | 		      std::swap( BaseClassRef::vind, tmp.BaseClassRef::vind );
1527 | 		      std::swap( BaseClassRef::m_leaf_max_size, tmp.BaseClassRef::m_leaf_max_size );
1528 | 		      std::swap( index_params, tmp.index_params );
1529 | 		      std::swap( treeIndex, tmp.treeIndex );
1530 | 		      std::swap( BaseClassRef::m_size, tmp.BaseClassRef::m_size );
1531 | 		      std::swap( BaseClassRef::m_size_at_index_build, tmp.BaseClassRef::m_size_at_index_build );
1532 | 		      std::swap( BaseClassRef::root_node, tmp.BaseClassRef::root_node );
1533 | 		      std::swap( BaseClassRef::root_bbox, tmp.BaseClassRef::root_bbox );
1534 | 		      std::swap( BaseClassRef::pool, tmp.BaseClassRef::pool );
1535 | 		      return *this;
1536 | 		 }
1537 | 
1538 | 		/**
1539 | 		 * Builds the index
1540 | 		 */
1541 | 		void buildIndex()
1542 | 		{
1543 | 			BaseClassRef::m_size = BaseClassRef::vind.size();
1544 | 			this->freeIndex(*this);
1545 | 			BaseClassRef::m_size_at_index_build = BaseClassRef::m_size;
1546 | 			if(BaseClassRef::m_size == 0) return;
1547 | 			computeBoundingBox(BaseClassRef::root_bbox);
1548 | 			BaseClassRef::root_node = this->divideTree(*this, 0, BaseClassRef::m_size, BaseClassRef::root_bbox );   // construct the tree
1549 | 		}
1550 | 
1551 | 		/** \name Query methods
1552 | 		  * @{ */
1553 | 
1554 | 		/**
1555 | 		 * Find set of nearest neighbors to vec[0:dim-1]. Their indices are stored inside
1556 | 		 * the result object.
1557 | 		 *
1558 | 		 * Params:
1559 | 		 *     result = the result object in which the indices of the nearest-neighbors are stored
1560 | 		 *     vec = the vector for which to search the nearest neighbors
1561 | 		 *
1562 | 		 * \tparam RESULTSET Should be any ResultSet<DistanceType>
1563 |          * \return  True if the requested neighbors could be found.
1564 | 		 * \sa knnSearch, radiusSearch
1565 | 		 */
1566 | 		template <typename RESULTSET>
1567 | 		bool findNeighbors(RESULTSET& result, const ElementType* vec, const SearchParams& searchParams) const
1568 | 		{
1569 | 			assert(vec);
1570 |             if (this->size(*this) == 0)
1571 |                 return false;
1572 | 			if (!BaseClassRef::root_node)
1573 |                 return false;
1574 | 			float epsError = 1 + searchParams.eps;
1575 | 
1576 | 			distance_vector_t dists; // fixed or variable-sized container (depending on DIM)
1577 | 			dists.assign((DIM > 0 ? DIM : BaseClassRef::dim) , 0); // Fill it with zeros.
1578 | 			DistanceType distsq = this->computeInitialDistances(*this, vec, dists);
1579 | 			searchLevel(result, vec, BaseClassRef::root_node, distsq, dists, epsError);  // "count_leaf" parameter removed since was neither used nor returned to the user.
1580 |             return result.full();
1581 | 		}
1582 | 
1583 | 		/**
1584 | 		 * Find the "num_closest" nearest neighbors to the \a query_point[0:dim-1]. Their indices are stored inside
1585 | 		 * the result object.
1586 | 		 *  \sa radiusSearch, findNeighbors
1587 | 		 * \note nChecks_IGNORED is ignored but kept for compatibility with the original FLANN interface.
1588 | 		 * \return Number `N` of valid points in the result set. Only the first `N` entries in `out_indices` and `out_distances_sq` will be valid. 
1589 | 		 *         Return may be less than `num_closest` only if the number of elements in the tree is less than `num_closest`.
1590 | 		 */
1591 | 		size_t knnSearch(const ElementType *query_point, const size_t num_closest, IndexType *out_indices, DistanceType *out_distances_sq, const int /* nChecks_IGNORED */ = 10) const
1592 | 		{
1593 | 			nanoflann::KNNResultSet<DistanceType,IndexType> resultSet(num_closest);
1594 | 			resultSet.init(out_indices, out_distances_sq);
1595 | 			this->findNeighbors(resultSet, query_point, nanoflann::SearchParams());
1596 | 			return resultSet.size();
1597 | 		}
1598 | 
1599 | 		/**
1600 | 		 * Find all the neighbors to \a query_point[0:dim-1] within a maximum radius.
1601 | 		 *  The output is given as a vector of pairs, of which the first element is a point index and the second the corresponding distance.
1602 | 		 *  Previous contents of \a IndicesDists are cleared.
1603 | 		 *
1604 | 		 *  If searchParams.sorted==true, the output list is sorted by ascending distances.
1605 | 		 *
1606 | 		 *  For a better performance, it is advisable to do a .reserve() on the vector if you have any wild guess about the number of expected matches.
1607 | 		 *
1608 | 		 *  \sa knnSearch, findNeighbors, radiusSearchCustomCallback
1609 | 		 * \return The number of points within the given radius (i.e. indices.size() or dists.size() )
1610 | 		 */
1611 | 		size_t radiusSearch(const ElementType *query_point, const DistanceType &radius, std::vector<std::pair<IndexType,DistanceType> >& IndicesDists, const SearchParams& searchParams) const
1612 | 		{
1613 | 			RadiusResultSet<DistanceType,IndexType> resultSet(radius, IndicesDists);
1614 | 			const size_t nFound = radiusSearchCustomCallback(query_point, resultSet, searchParams);
1615 | 			if (searchParams.sorted)
1616 | 				std::sort(IndicesDists.begin(), IndicesDists.end(), IndexDist_Sorter() );
1617 | 			return nFound;
1618 | 		}
1619 | 
1620 | 		/**
1621 | 		 * Just like radiusSearch() but with a custom callback class for each point found in the radius of the query.
1622 | 		 * See the source of RadiusResultSet<> as a start point for your own classes.
1623 | 		 * \sa radiusSearch
1624 | 		 */
1625 | 		template <class SEARCH_CALLBACK>
1626 | 		size_t radiusSearchCustomCallback(const ElementType *query_point, SEARCH_CALLBACK &resultSet, const SearchParams& searchParams = SearchParams() ) const
1627 | 		{
1628 | 			this->findNeighbors(resultSet, query_point, searchParams);
1629 | 			return resultSet.size();
1630 | 		}
1631 | 
1632 | 		/** @} */
1633 | 
1634 | 	public:
1635 | 
1636 | 
1637 | 		void computeBoundingBox(BoundingBox& bbox)
1638 | 		{
1639 | 			bbox.resize((DIM > 0 ? DIM : BaseClassRef::dim));
1640 | 			if (dataset.kdtree_get_bbox(bbox))
1641 | 			{
1642 | 				// Done! It was implemented in derived class
1643 | 			}
1644 | 			else
1645 | 			{
1646 | 				const size_t N = BaseClassRef::m_size;
1647 | 				if (!N) throw std::runtime_error("[nanoflann] computeBoundingBox() called but no data points found.");
1648 | 				for (int i = 0; i < (DIM > 0 ? DIM : BaseClassRef::dim); ++i) {
1649 | 					bbox[i].low =
1650 | 					bbox[i].high = this->dataset_get(*this, BaseClassRef::vind[0], i);
1651 | 				}
1652 | 				for (size_t k = 1; k < N; ++k) {
1653 | 					for (int i = 0; i < (DIM > 0 ? DIM : BaseClassRef::dim); ++i) {
1654 | 						if (this->dataset_get(*this, BaseClassRef::vind[k], i) < bbox[i].low) bbox[i].low = this->dataset_get(*this, BaseClassRef::vind[k], i);
1655 | 						if (this->dataset_get(*this, BaseClassRef::vind[k], i) > bbox[i].high) bbox[i].high = this->dataset_get(*this, BaseClassRef::vind[k], i);
1656 | 					}
1657 | 				}
1658 | 			}
1659 | 		}
1660 | 
1661 | 		/**
1662 | 		 * Performs an exact search in the tree starting from a node.
1663 | 		 * \tparam RESULTSET Should be any ResultSet<DistanceType>
1664 | 		 */
1665 | 		template <class RESULTSET>
1666 | 		void searchLevel(RESULTSET& result_set, const ElementType* vec, const NodePtr node, DistanceType mindistsq,
1667 | 						 distance_vector_t& dists, const float epsError) const
1668 | 		{
1669 | 			/* If this is a leaf node, then do check and return. */
1670 | 			if ((node->child1 == NULL) && (node->child2 == NULL)) {
1671 | 				//count_leaf += (node->lr.right-node->lr.left);  // Removed since was neither used nor returned to the user.
1672 | 				DistanceType worst_dist = result_set.worstDist();
1673 | 				for (IndexType i = node->node_type.lr.left; i < node->node_type.lr.right; ++i) {
1674 | 					const IndexType index = BaseClassRef::vind[i];// reorder... : i;
1675 | 					if(treeIndex[index] == -1)
1676 | 						continue;
1677 | 					DistanceType dist = distance.evalMetric(vec, index, (DIM > 0 ? DIM : BaseClassRef::dim));
1678 | 					if (dist<worst_dist) {
1679 | 						result_set.addPoint(dist, BaseClassRef::vind[i]);
1680 | 					}
1681 | 				}
1682 | 				return;
1683 | 			}
1684 | 
1685 | 			/* Which child branch should be taken first? */
1686 | 			int idx = node->node_type.sub.divfeat;
1687 | 			ElementType val = vec[idx];
1688 | 			DistanceType diff1 = val - node->node_type.sub.divlow;
1689 | 			DistanceType diff2 = val - node->node_type.sub.divhigh;
1690 | 
1691 | 			NodePtr bestChild;
1692 | 			NodePtr otherChild;
1693 | 			DistanceType cut_dist;
1694 | 			if ((diff1 + diff2) < 0) {
1695 | 				bestChild = node->child1;
1696 | 				otherChild = node->child2;
1697 | 				cut_dist = distance.accum_dist(val, node->node_type.sub.divhigh, idx);
1698 | 			}
1699 | 			else {
1700 | 				bestChild = node->child2;
1701 | 				otherChild = node->child1;
1702 | 				cut_dist = distance.accum_dist( val, node->node_type.sub.divlow, idx);
1703 | 			}
1704 | 
1705 | 			/* Call recursively to search next level down. */
1706 | 			searchLevel(result_set, vec, bestChild, mindistsq, dists, epsError);
1707 | 
1708 | 			DistanceType dst = dists[idx];
1709 | 			mindistsq = mindistsq + cut_dist - dst;
1710 | 			dists[idx] = cut_dist;
1711 | 			if (mindistsq*epsError <= result_set.worstDist()) {
1712 | 				searchLevel(result_set, vec, otherChild, mindistsq, dists, epsError);
1713 | 			}
1714 | 			dists[idx] = dst;
1715 | 		}
1716 | 
1717 | 	public:
1718 | 		/**  Stores the index in a binary file.
1719 | 		  *   IMPORTANT NOTE: The set of data points is NOT stored in the file, so when loading the index object it must be constructed associated to the same source of data points used while building it.
1720 | 		  * See the example: examples/saveload_example.cpp
1721 | 		  * \sa loadIndex  */
1722 | 		void saveIndex(FILE* stream)
1723 | 		{
1724 | 			this->saveIndex_(*this, stream);
1725 | 		}
1726 | 
1727 | 		/**  Loads a previous index from a binary file.
1728 | 		  *   IMPORTANT NOTE: The set of data points is NOT stored in the file, so the index object must be constructed associated to the same source of data points used while building the index.
1729 | 		  * See the example: examples/saveload_example.cpp
1730 | 		  * \sa loadIndex  */
1731 | 		void loadIndex(FILE* stream)
1732 | 		{
1733 | 			this->loadIndex_(*this, stream);
1734 | 		}
1735 | 
1736 | 	};
1737 | 
1738 | 
1739 | 	/** kd-tree dynaimic index
1740 | 	 *
1741 | 	 * class to create multiple static index and merge their results to behave as single dynamic index as proposed in Logarithmic Approach.
1742 | 	 *  
1743 | 	 *  Example of usage:
1744 | 	 *  examples/dynamic_pointcloud_example.cpp
1745 | 	 *
1746 | 	 * \tparam DatasetAdaptor The user-provided adaptor (see comments above).
1747 | 	 * \tparam Distance The distance metric to use: nanoflann::metric_L1, nanoflann::metric_L2, nanoflann::metric_L2_Simple, etc.
1748 | 	 * \tparam DIM Dimensionality of data points (e.g. 3 for 3D points)
1749 | 	 * \tparam IndexType Will be typically size_t or int
1750 | 	 */
1751 | 	template <typename Distance, class DatasetAdaptor,int DIM = -1, typename IndexType = size_t>
1752 | 	class KDTreeSingleIndexDynamicAdaptor
1753 | 	{
1754 | 	public:
1755 | 		typedef typename Distance::ElementType  ElementType;
1756 | 		typedef typename Distance::DistanceType DistanceType;
1757 | 	protected:
1758 | 
1759 | 		size_t m_leaf_max_size;
1760 | 		size_t treeCount;
1761 | 		size_t pointCount;
1762 | 
1763 | 		/**
1764 | 		 * The dataset used by this index
1765 | 		 */
1766 | 		DatasetAdaptor &dataset; //!< The source of our data
1767 | 
1768 | 		std::vector<int> treeIndex; //!< treeIndex[idx] is the index of tree in which point at idx is stored. treeIndex[idx]=-1 means that point has been removed.
1769 | 
1770 | 		KDTreeSingleIndexAdaptorParams index_params;
1771 | 
1772 | 		int dim;  //!< Dimensionality of each data point
1773 | 
1774 | 		typedef KDTreeSingleIndexDynamicAdaptor_<Distance, DatasetAdaptor, DIM> index_container_t;
1775 | 		std::vector<index_container_t> index;
1776 | 
1777 | 	public:
1778 | 		/** Get a const ref to the internal list of indices; the number of indices is adapted dynamically as 
1779 | 		  * the dataset grows in size. */
1780 | 		const std::vector<index_container_t> & getAllIndices() const {
1781 | 			return index;
1782 | 		}
1783 | 
1784 | 	private:
1785 | 		/** finds position of least significant unset bit */
1786 | 		int First0Bit(IndexType num)
1787 | 		{
1788 | 			int pos = 0;
1789 | 			while(num&1)
1790 | 			{
1791 | 				num = num>>1;
1792 | 				pos++;
1793 | 			}
1794 | 			return pos;
1795 | 		}
1796 | 
1797 | 		/** Creates multiple empty trees to handle dynamic support */
1798 | 		void init()
1799 | 		{
1800 | 			typedef KDTreeSingleIndexDynamicAdaptor_<Distance, DatasetAdaptor, DIM> my_kd_tree_t;
1801 | 			std::vector<my_kd_tree_t> index_(treeCount, my_kd_tree_t(dim /*dim*/, dataset, treeIndex, index_params));
1802 | 			index=index_;
1803 | 		}
1804 | 
1805 | 	public:
1806 | 
1807 | 		Distance distance;
1808 | 
1809 | 		/**
1810 | 		 * KDTree constructor
1811 | 		 *
1812 | 		 * Refer to docs in README.md or online in https://github.com/jlblancoc/nanoflann
1813 | 		 *
1814 | 		 * The KD-Tree point dimension (the length of each point in the datase, e.g. 3 for 3D points)
1815 | 		 * is determined by means of:
1816 | 		 *  - The \a DIM template parameter if >0 (highest priority)
1817 | 		 *  - Otherwise, the \a dimensionality parameter of this constructor.
1818 | 		 *
1819 | 		 * @param inputData Dataset with the input features
1820 | 		 * @param params Basically, the maximum leaf node size
1821 | 		 */
1822 | 		KDTreeSingleIndexDynamicAdaptor(const int dimensionality, DatasetAdaptor& inputData, const KDTreeSingleIndexAdaptorParams& params = KDTreeSingleIndexAdaptorParams() , const size_t maximumPointCount = 1000000000U) :
1823 | 			dataset(inputData), index_params(params), distance(inputData)
1824 | 		{
1825 | 			if (dataset.kdtree_get_point_count()) throw std::runtime_error("[nanoflann] cannot handle non empty point cloud.");
1826 | 			treeCount = log2(maximumPointCount);
1827 | 			pointCount = 0U;
1828 | 			dim = dimensionality;
1829 | 			treeIndex.clear();
1830 | 			if (DIM > 0) dim = DIM;
1831 | 			m_leaf_max_size = params.leaf_max_size;
1832 | 			init();
1833 | 		}
1834 | 
1835 | 
1836 | 		/** Add points to the set, Inserts all points from [start, end] */
1837 | 		void addPoints(IndexType start, IndexType end)
1838 | 		{
1839 | 			int count = end - start + 1;
1840 | 			treeIndex.resize(treeIndex.size() + count);
1841 | 			for(IndexType idx = start; idx <= end; idx++) {
1842 | 				int pos = First0Bit(pointCount);
1843 | 				index[pos].vind.clear();
1844 | 				treeIndex[pointCount]=pos;
1845 | 				for(int i = 0; i < pos; i++) {
1846 | 					for(int j = 0; j < static_cast<int>(index[i].vind.size()); j++) {
1847 | 						index[pos].vind.push_back(index[i].vind[j]);
1848 | 						treeIndex[index[i].vind[j]] = pos;
1849 | 					}
1850 | 					index[i].vind.clear();
1851 | 					index[i].freeIndex(index[i]);
1852 | 				}
1853 | 				index[pos].vind.push_back(idx);
1854 | 				index[pos].buildIndex();
1855 | 				pointCount++;
1856 | 			}
1857 | 		}
1858 | 
1859 | 		/** Remove a point from the set (Lazy Deletion) */
1860 | 		void removePoint(size_t idx)
1861 | 		{
1862 | 			if(idx >= pointCount)
1863 | 				return;
1864 | 			treeIndex[idx] = -1;
1865 | 		}
1866 | 
1867 | 		/**
1868 | 		 * Find set of nearest neighbors to vec[0:dim-1]. Their indices are stored inside
1869 | 		 * the result object.
1870 | 		 *
1871 | 		 * Params:
1872 | 		 *     result = the result object in which the indices of the nearest-neighbors are stored
1873 | 		 *     vec = the vector for which to search the nearest neighbors
1874 | 		 *
1875 | 		 * \tparam RESULTSET Should be any ResultSet<DistanceType>
1876 |          * \return  True if the requested neighbors could be found.
1877 | 		 * \sa knnSearch, radiusSearch
1878 | 		 */
1879 | 		template <typename RESULTSET>
1880 | 		bool findNeighbors(RESULTSET& result, const ElementType* vec, const SearchParams& searchParams) const
1881 | 		{
1882 | 			for(size_t i = 0; i < treeCount; i++)
1883 | 			{
1884 | 				index[i].findNeighbors(result, &vec[0], searchParams);
1885 | 			}
1886 | 			return result.full();
1887 | 		}
1888 | 
1889 | 	}; 
1890 | 
1891 | 	/** An L2-metric KD-tree adaptor for working with data directly stored in an Eigen Matrix, without duplicating the data storage.
1892 | 	  *  Each row in the matrix represents a point in the state space.
1893 | 	  *
1894 | 	  *  Example of usage:
1895 | 	  * \code
1896 | 	  * 	Eigen::Matrix<num_t,Dynamic,Dynamic>  mat;
1897 | 	  * 	// Fill out "mat"...
1898 | 	  *
1899 | 	  * 	typedef KDTreeEigenMatrixAdaptor< Eigen::Matrix<num_t,Dynamic,Dynamic> >  my_kd_tree_t;
1900 | 	  * 	const int max_leaf = 10;
1901 | 	  * 	my_kd_tree_t   mat_index(mat, max_leaf );
1902 | 	  * 	mat_index.index->buildIndex();
1903 | 	  * 	mat_index.index->...
1904 | 	  * \endcode
1905 | 	  *
1906 | 	  *  \tparam DIM If set to >0, it specifies a compile-time fixed dimensionality for the points in the data set, allowing more compiler optimizations.
1907 | 	  *  \tparam Distance The distance metric to use: nanoflann::metric_L1, nanoflann::metric_L2, nanoflann::metric_L2_Simple, etc.
1908 | 	  */
1909 | 	template <class MatrixType, class Distance = nanoflann::metric_L2>
1910 | 	struct KDTreeEigenMatrixAdaptor
1911 | 	{
1912 | 		typedef KDTreeEigenMatrixAdaptor<MatrixType,Distance> self_t;
1913 | 		typedef typename MatrixType::Scalar              num_t;
1914 | 		typedef typename MatrixType::Index IndexType;
1915 | 		typedef typename Distance::template traits<num_t,self_t>::distance_t metric_t;
1916 | 		typedef KDTreeSingleIndexAdaptor< metric_t,self_t, MatrixType::ColsAtCompileTime,IndexType>  index_t;
1917 | 
1918 | 		index_t* index; //! The kd-tree index for the user to call its methods as usual with any other FLANN index.
1919 | 
1920 | 		/// Constructor: takes a const ref to the matrix object with the data points
1921 | 		KDTreeEigenMatrixAdaptor(const MatrixType &mat, const int leaf_max_size = 10) : m_data_matrix(mat)
1922 | 		{
1923 | 			const IndexType dims = mat.cols();
1924 | 			index = new index_t( dims, *this /* adaptor */, nanoflann::KDTreeSingleIndexAdaptorParams(leaf_max_size ) );
1925 | 			index->buildIndex();
1926 | 		}
1927 | 	private:
1928 | 		/** Hidden copy constructor, to disallow copying this class (Not implemented) */
1929 | 		KDTreeEigenMatrixAdaptor(const self_t&);
1930 | 	public:
1931 | 
1932 | 		~KDTreeEigenMatrixAdaptor() {
1933 | 			delete index;
1934 | 		}
1935 | 
1936 | 		const MatrixType &m_data_matrix;
1937 | 
1938 | 		/** Query for the \a num_closest closest points to a given point (entered as query_point[0:dim-1]).
1939 | 		  *  Note that this is a short-cut method for index->findNeighbors().
1940 | 		  *  The user can also call index->... methods as desired.
1941 | 		  * \note nChecks_IGNORED is ignored but kept for compatibility with the original FLANN interface.
1942 | 		  */
1943 | 		inline void query(const num_t *query_point, const size_t num_closest, IndexType *out_indices, num_t *out_distances_sq, const int /* nChecks_IGNORED */ = 10) const
1944 | 		{
1945 | 			nanoflann::KNNResultSet<num_t, IndexType> resultSet(num_closest);
1946 | 			resultSet.init(out_indices, out_distances_sq);
1947 | 			index->findNeighbors(resultSet, query_point, nanoflann::SearchParams());
1948 | 		}
1949 | 
1950 | 		/** @name Interface expected by KDTreeSingleIndexAdaptor
1951 | 		  * @{ */
1952 | 
1953 | 		const self_t & derived() const {
1954 | 			return *this;
1955 | 		}
1956 | 		self_t & derived()       {
1957 | 			return *this;
1958 | 		}
1959 | 
1960 | 		// Must return the number of data points
1961 | 		inline size_t kdtree_get_point_count() const {
1962 | 			return m_data_matrix.rows();
1963 | 		}
1964 | 
1965 | 		// Returns the dim'th component of the idx'th point in the class:
1966 | 		inline num_t kdtree_get_pt(const IndexType idx, int dim) const {
1967 | 			return m_data_matrix.coeff(idx, IndexType(dim));
1968 | 		}
1969 | 
1970 | 		// Optional bounding-box computation: return false to default to a standard bbox computation loop.
1971 | 		//   Return true if the BBOX was already computed by the class and returned in "bb" so it can be avoided to redo it again.
1972 | 		//   Look at bb.size() to find out the expected dimensionality (e.g. 2 or 3 for point clouds)
1973 | 		template <class BBOX>
1974 | 		bool kdtree_get_bbox(BBOX& /*bb*/) const {
1975 | 			return false;
1976 | 		}
1977 | 
1978 | 		/** @} */
1979 | 
1980 | 	}; // end of KDTreeEigenMatrixAdaptor
1981 | 	/** @} */
1982 | 
1983 | /** @} */ // end of grouping
1984 | } // end of NS
1985 | 
1986 | 
1987 | #endif /* NANOFLANN_HPP_ */
1988 | 


--------------------------------------------------------------------------------
/include/point.h:
--------------------------------------------------------------------------------
 1 | #ifndef POINT_H
 2 | #define POINT_H
 3 | 
 4 | #include <string>
 5 | 
 6 | class point_t {
 7 | public:
 8 |     /*
 9 |      * constructors
10 |      */
11 |     point_t();
12 |     point_t(float x, float y, float z);
13 | 
14 |     // unpack values into a float array
15 |     void unpack(float * array);
16 | 
17 |     // get element of point by index. used for nanoflann kd tree
18 |     float get_element(int index);
19 | 
20 |     // pretty print
21 |     std::string to_string();
22 | 
23 | private:
24 |     float elems[3];
25 | };
26 | 
27 | #endif
28 | 


--------------------------------------------------------------------------------
/include/warp_field.h:
--------------------------------------------------------------------------------
 1 | #ifndef WARP_FIELD_H
 2 | #define WARP_FIELD_H
 3 | 
 4 | #include <vector>
 5 | #include "dual_quat.h"
 6 | #include "deform_node.h"
 7 | #include "frame.h"
 8 | 
 9 | // includes for nanoflann, as well as a typedef for brevity
10 | #include "nanoflann.hpp"
11 | using namespace nanoflann;
12 | 
13 | class warp_field_t;
14 | typedef KDTreeSingleIndexAdaptor<
15 |     L2_Simple_Adaptor<float, warp_field_t>, 
16 |     warp_field_t, 
17 |     3
18 | > kd_tree_t;
19 | 
20 | class warp_field_t {
21 | public:
22 |     // constructor
23 |     warp_field_t();
24 | 
25 |     // destructor
26 |     ~warp_field_t();
27 |     
28 |     /*
29 |      *  interface required for nanoflann implementation of kd-tree
30 |      */
31 |     // size of the warp field
32 |     size_t kdtree_get_point_count() const;
33 |     
34 |     // access individual elements of points in cloud
35 |     float kdtree_get_pt(const size_t idx, int dim) const;
36 | 
37 |     // unused, optional interface method to determine bounding box
38 |     template <class BBOX>
39 |     bool kdtree_get_bbox(BBOX& bb) const {
40 |         return false;
41 |     }
42 |     
43 |     // retrieve the N closest neighbouring nodes from a given point
44 |     std::vector<deform_node_t *> find_neighbours(point_t p);
45 | 
46 | private:
47 |     // constants
48 |     const static int NUM_NEIGHBOURS = 8;
49 | 
50 |     // kd-tree for efficient closest neighbour search
51 |     kd_tree_t * kd_tree;
52 | 
53 |     // vector of deformation nodes in the warp field defining the state of the warp field
54 |     std::vector<deform_node_t *> nodes;
55 | 
56 |     /*
57 |      *  private methods 
58 |      */
59 |     // calculate dual-quaternion blending for a given point
60 |     dual_quat_t dual_quaternion_blending(point_t p);    
61 | 
62 |     // updates the warp field
63 |     void update_warp_field();
64 | };
65 | 
66 | #endif
67 | 


--------------------------------------------------------------------------------
/installDependencies.sh:
--------------------------------------------------------------------------------
 1 | # CMake
 2 | sudo apt-get install cmake
 3 | 
 4 | #
 5 | # prerequisites for ceres
 6 | #
 7 | 
 8 | # google-glog + gflags
 9 | sudo apt-get install libgoogle-glog-dev
10 | # BLAS & LAPACK
11 | sudo apt-get install libatlas-base-dev
12 | # Eigen3
13 | sudo apt-get install libeigen3-dev
14 | # SuiteSparse and CXSparse (optional)
15 | # - If you want to build Ceres as a *static* library (the default)
16 | #   you can use the SuiteSparse package in the main Ubuntu package
17 | #   repository:
18 | sudo apt-get install libsuitesparse-dev
19 | sudo apt-get update
20 | sudo apt-get install libsuitesparse-dev
21 | 
22 | # 
23 | # ceres solver
24 | #
25 | if cd build/ceres-solver; then
26 |     git pull
27 |     cd ..
28 | else
29 |     cd build 
30 |     git clone https://github.com/ceres-solver/ceres-solver.git
31 |     cd ceres-solver
32 |     cd build
33 |     cmake ..
34 |     make -j3
35 |     sudo make install
36 |     cd ../..
37 | fi
38 | cd ..
39 | 
40 | 


--------------------------------------------------------------------------------
/projectBuild.sh:
--------------------------------------------------------------------------------
1 | cd build
2 | cmake CMakeLists.txt .
3 | make
4 | ./dynamicfusion
5 | 


--------------------------------------------------------------------------------
/report/build.sh:
--------------------------------------------------------------------------------
1 | rm report.pdf
2 | latex report
3 | bibtex report
4 | latex report
5 | latex report
6 | evince report.pdf
7 | 


--------------------------------------------------------------------------------
/report/report.bib:
--------------------------------------------------------------------------------
 1 | @INPROCEEDINGS{kinectfusion,
 2 | author={R. A. Newcombe and S. Izadi and O. Hilliges and D. Molyneaux and D. Kim and A. J. Davison and P. Kohi and J. Shotton and S. Hodges and A. Fitzgibbon},
 3 | booktitle={2011 10th IEEE International Symposium on Mixed and Augmented Reality},
 4 | title={KinectFusion: Real-time dense surface mapping and tracking},
 5 | year={2011},
 6 | volume={},
 7 | number={},
 8 | pages={127-136},
 9 | keywords={Cameras;Image reconstruction;Iterative closest point algorithm;Real time systems;Simultaneous localization and mapping;Surface reconstruction;Three dimensional displays;AR;Dense Reconstruction;Depth Cameras;GPU;Real-Time;SLAM;Tracking;Volumetric Representation},
10 | doi={10.1109/ISMAR.2011.6092378},
11 | ISSN={},
12 | month={Oct},}
13 | volume deform
14 | killing fusion
15 | 
16 | @INPROCEEDINGS{dynamicfusion,
17 | author={R. A. Newcombe and D. Fox and S. M. Seitz},
18 | booktitle={2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
19 | title={DynamicFusion: Reconstruction and tracking of non-rigid scenes in real-time},
20 | year={2015},
21 | volume={},
22 | number={},
23 | pages={343-352},
24 | keywords={geometry;image denoising;image reconstruction;motion estimation;object tracking;DynamicFusion approach;KinectFusion;RGBD scans;SLAM system;commodity sensors;denoised reconstructions;dense volumetric 6D motion field estimation;detailed reconstructions;live frame;nonrigid scenes tracking;nonrigidly deforming scenes reconstruction;scene geometry reconstruction;Cameras;Geometry;Real-time systems;Shape;Surface reconstruction;Tracking;Transforms},
25 | doi={10.1109/CVPR.2015.7298631},
26 | ISSN={1063-6919},
27 | month={June}
28 | }
29 | 
30 | @article{kdtree,
31 |  author = {Bentley, Jon Louis},
32 |  title = {Multidimensional Binary Search Trees Used for Associative Searching},
33 |  journal = {Commun. ACM},
34 |  issue_date = {Sept. 1975},
35 |  volume = {18},
36 |  number = {9},
37 |  month = sep,
38 |  year = {1975},
39 |  issn = {0001-0782},
40 |  pages = {509--517},
41 |  numpages = {9},
42 |  url = {http://doi.acm.org/10.1145/361002.361007},
43 |  doi = {10.1145/361002.361007},
44 |  acmid = {361007},
45 |  publisher = {ACM},
46 |  address = {New York, NY, USA},
47 |  keywords = {associative retrieval, attribute, binary search trees, binary tree insertion, information retrieval system, intersection queries, key, nearest neighbor queries, partial match queries},
48 | } 
49 | 
50 | 
51 | @article{icp,
52 |     author = "Paul J. Besl and Neil D. McKay",
53 |     title = "A Method for Registration of 3-D Shapes",
54 |     journal = "IEEE Transactions on Pattern Analysis and Machine Intelligence",
55 |     volume = "14",
56 |     number = "2",
57 |     year = "1992"
58 | }
59 | 
60 | 
61 | 
62 | @article{killingfusion,
63 | author = {Slavcheva, Miroslava and Baust, Maximilian and Cremers, Daniel and Ilic, Slobodan},
64 | year = {2017},
65 | month = {07},
66 | pages = {},
67 | title = {KillingFusion: Non-rigid 3D Reconstruction without Correspondences}
68 | }
69 | 
70 | @inproceedings{volumedeform,
71 |   title={VolumeDeform: Real-Time Volumetric Non-rigid Reconstruction},
72 |   author={Matthias Innmann and Michael Zollhofer and Matthias Niessner and Christian Theobalt and Marc Stamminger},
73 |   booktitle={ECCV},
74 |   year={2016}
75 | }
76 | 


--------------------------------------------------------------------------------
/report/report.tex:
--------------------------------------------------------------------------------
  1 | %&pdflatex
  2 | \documentclass[a4paper]{article}
  3 | 
  4 | %% Language and font encodings
  5 | \usepackage[english]{babel}
  6 | \usepackage[utf8x]{inputenc}
  7 | \usepackage[T1]{fontenc}
  8 | 
  9 | %% Sets page size and margins
 10 | \usepackage[a4paper,top=3cm,bottom=2cm,left=3cm,right=3cm,marginparwidth=1.75cm]{geometry}
 11 | 
 12 | %% Useful packages
 13 | \usepackage{amsmath}
 14 | \usepackage{graphicx}
 15 | \usepackage{listings}
 16 | \usepackage[colorinlistoftodos]{todonotes}
 17 | \usepackage[colorlinks=true, allcolors=blue]{hyperref}
 18 | 
 19 | \title{Real-time reconstruction of non-rigid objects from RGBD data: Interim Report}
 20 | \author{Amelia Gordafarid Crowther}
 21 | 
 22 | \begin{document}
 23 | 
 24 | \maketitle
 25 | \section{Introduction}
 26 | 
 27 | \subsection{Motivation}
 28 | 
 29 | The reconstruction of non-rigid objects is a popular area of research currently, with many useful applications in fields such as Robotics, Medicine and Augmented Reality. 
 30 | 
 31 | However, most existing systems are designed only to reconstruct static geometry. This leads to hard limitations in the adaptability of the system to a changing environment. Worse still, highly deformable or dynamic objects such as a moving person cannot be reconstructed at all by a static system.
 32 | 
 33 | There has been a lot of interest in techniques for non-rigid reconstruction, but despite it being a popular research topic, few  implementations of these proposed systems are widely available. 
 34 | 
 35 | The aim of this project is to provide such an implementation that can reconstruct a rigid object from a  non-rigid scene in real-time.
 36 | 
 37 | 
 38 | \section{Background}
 39 | 
 40 | \subsection{Kinect Camera}
 41 | 
 42 | The first and most fundamental element of the system is the sensor used to obtain information about the surface to be reconstructed. The most commonly used sensor for this purpose is the Kinect camera. Originally designed as a commodity level camera for home games consoles, it was available directly to consumers. This means that Kinect cameras are widely and cheaply available to use.
 43 | 
 44 | However, they provide high quality information considering their price. This is provided to the system in the form of an ordinary RGB image, augmented with a fourth parameter indicating the perceived depth from the camera at that pixel. Using the available depth information, it is also possible to compute the normal to the surface at each pixel of the image. In some cases, depth information may be missing at certain pixels, producing so-called `holes' in the depth image. Errors such as these are not difficult to overcome but must be handled.
 45 | 
 46 | 
 47 | The sensor is capable of producing an RGB-D image to the system in 640x480 resolution at up to 30Hz. This is important, as it determines the amount of information that can be added to the system at each update. Additionally, it places a hard limit on the update rate of the system, which can never exceed the sensor rate.
 48 | 
 49 | \subsection{Point Cloud Correspondences}
 50 | 
 51 | Once all noise has been removed from the RGB-D image, it can be converted into a point cloud, which forms the input to the system at each update. Every point in the cloud is considered to lie on the surface of the object that is being reconstructed. 
 52 | 
 53 | The main task of the system, which has been approached using multiple methods, is to find correspondences between the point clouds obtained at each update. These correspondences can be used to resolve all the data obtained into one model, which is the target surface for reconstruction.
 54 | 
 55 | 
 56 | \subsection{KinectFusion}
 57 | 
 58 | KinectFusion\cite{kinectfusion} provides one such method for determining correspondences between point clouds. It makes one very important assumption: that the surface it is reconstructing is rigid, and will not be deformed while it is being observed by the system. While this is a very restricted approach to the problem, this algorithm laid the groundwork for later non-rigid systems such as DynamicFusion.
 59 | 
 60 | \subsubsection{Iterative Closest Point algorithm}
 61 | Iterative Closest Point\cite{icp} (ICP) is a widely used algorithm in this field which has been optimised and adapted for many purposes. Its purpose is simply to take two point clouds and return the transformation which most closely aligns the two clouds.
 62 | Psuedo-code for the ICP algorithm is shown below:
 63 | 
 64 | \begin{lstlisting}[language=Python]
 65 | def ICP(P, Q):
 66 |     transform = identity
 67 |     while not aligned(P, Q):
 68 |         correspondence = []
 69 |         for p in P:
 70 |             p_t = apply(transform, p)
 71 |             q = closest point to p_t in Q
 72 |             correspondence += (p, q)
 73 |         transform = least_squares(correspondence)
 74 |     return transform
 75 | \end{lstlisting}
 76 | 
 77 | \subsubsection{Principles of operation}
 78 | 
 79 | Initially, a pre-processing phase occurs, where depth and normal maps are constructed from the input data. The system then enters a continuously updating loop with three stages.
 80 | 
 81 | \begin{enumerate}
 82 | \item \textbf{Sensor pose estimation}: KinectFusion uses the ICP algorithm to align its prediction of the surface with the actual measurements of the surface it receives. The transform that is returned is rigid, which functions perfectly for this algorithm, but requires adaptation for a non-rigid case. 
 83 | 
 84 | \item \textbf{Update surface reconstruction}: Here, the system receives input from the sensor, which it uses to create a truncated signed distance function (TSDF) for that live frame. Together with the TSDFs from previous frames, a composite TSDF is created to represent the system's model of the scene. This composite TSDF is discretised into a 2-dimensional array and stored in global GPU memory.
 85 | 
 86 | \item \textbf{Surface prediction}: At this stage KinectFusion will produce its own depth map by ray-casting against its own model of the scene. This produces a prediction of what it expects the scene to look like. 
 87 | \end{enumerate}
 88 | 
 89 | 
 90 | \subsubsection{Limitations}
 91 | 
 92 | The main limitation of the KinectFusion system is that it fails to handle deforming geometry. If a surface is not completely static then it will fail to reconstruct it.  
 93 | 
 94 | 
 95 | There have been several methods proposed for the reconstruction of non-rigid surfaces. DynamicFusion is the most 
 96 | 
 97 | \newpage
 98 | \subsection{DynamicFusion}
 99 | 
100 | DynamicFusion\cite{dynamicfusion} was the first system to achieve dense, dynamic scene reconstruction in real-time. Both VolumeDeform and KillingFusion were heavily inspired by or based on DynamicFusion.
101 | 
102 | \subsubsection{Warp field}
103 | 
104 | The warp field is essentially a set of deformation nodes, each of which is denoted:
105 | 
106 | $$\mathcal{N}^t_{\textbf{warp}}=\{ \textbf{dg}_v, \textbf{dg}_w, \textbf{dg}_{se3} \}$$
107 | 
108 | Where the parameters refer to:
109 | 
110 | \begin{itemize}
111 | \item $\textbf{dg}_v$: The position vector of the deformation node
112 | 
113 | \item $\textbf{dg}_w$: The radius of influence of the node, which determines the strength of the application of the transformation to the node's neighbours
114 | 
115 | \item $\textbf{dg}_{se3}$: The transformation of the deformation node, represented using dual quaternions
116 | \end{itemize}
117 | 
118 | 
119 | \subsubsection{Dual quaternion blending}
120 | 
121 | \textbf{Dual quaternions}
122 | 
123 | Dual quaternions consist of eight elements providing six degrees of freedom. They are split across two quaternions, named the real and dual parts respectively.  They can be used to represent any transformation consisting of a translation and a rotation around an axis.\\
124 | 
125 | \noindent\textbf{KD-Tree}
126 | 
127 | 
128 | KD-Trees\cite{kdtree} are data structures which are optimised to efficiently find the nearest neighbour to a given point, as well as all points within a range. They are very useful when performing dual quaternion blending, as they allow us to restrict the problem space to only the local deformation nodes on the surface.\\
129 | 
130 | Using these definitions, we are able to define dual quaternion blending. Intuitively, it is the average of the local dual quaternions of the deformation nodes, each weighted according to their radius of influence. Here is the definition from DynamicFusion:
131 | 
132 | 
133 | $$\textbf{DQB}(x_c) = \frac{\sum_{k \in N(x_c)} \textbf{w}_k(x_c)\hat{\textbf{q}}_{kc}}{\Vert \sum_{k \in N(x_c)} \textbf{w}_k(x_c)\hat{\textbf{q}}_{kc} \Vert}$$
134 | 
135 | In this context, $N(x_c)$ is the set of the $k$ nearest deformation nodes. These can be found very efficciently using a KD-Tree.
136 | 
137 | $\textbf{w}_k(x_c)$ is the weight of the $k^{th}$ deformation node, which can be derived using the following formula:
138 | $$\textbf{w}_i(x_c) = \exp \Big(\frac{-\Vert \textbf{dg}^i_v - x_c \Vert ^2}{2(\textbf{dg}^i_w)^2}\Big)$$ 
139 | 
140 | \subsubsection{Warp function}
141 | 
142 | Once dual quaternion blending has been applied, there is only one more step before we can produce the final warp function. The result must be converted from a quaternion back into a transformation matrix.
143 | 
144 | Now we can define the warp function:
145 | 
146 | $$\mathcal{W}(x_c) = SE3(\textbf{DQB}(x_c))$$
147 | 
148 | This function is capable of warping any deformation node in the canonical model into the live frame.
149 | 
150 | \subsubsection{Principles of operation}
151 | 
152 | 
153 | There are three key steps that occur in each iteration of the DynamicFusion algorithm, and they each take place whenever a new live frame is produced:
154 | 
155 | \begin{enumerate}
156 | \item \textbf{Warp field estimation}: The warp field is estimated by choosing parameters such that the error function below is minimised:
157 | 
158 | $$E(\mathcal{W}_t, \mathcal{V}, D_t, \mathcal{E}) = \textbf{Data}(\mathcal{W}_t, \mathcal{V}_t, D_t) + \lambda\textbf{Reg}(\mathcal{W}_t, \mathcal{E})$$ 
159 | 
160 | This function consists of two terms:
161 | \begin{itemize}
162 | \item $\textbf{Data}(\mathcal{W}_t, \mathcal{V}_t, D_t) = \sum\limits_{u \in \Omega} \psi_{data}(\hat{\textbf{n}}_u^\top (\hat{\textbf{v}}_u - \textbf{vl}_u))$ 
163 | 
164 | This term represents the ICP cost between the model and the live frame. This step is similar to KinectFusion, although it uses a non-rigid version of ICP.
165 | 
166 | \item $\textbf{Reg}(\mathcal{W}_t, \mathcal{E}) = \sum\limits_{i=0}^n\sum\limits_{j \in \mathcal{E}(i)}\alpha_{ij}\psi_{reg}(\textbf{T}_{ic}\textbf{dg}^j_v - \textbf{T}_{jc}\textbf{dg}^j_v)$ 
167 | 
168 | This term which regularises the warp field. This ensures that motion is as smooth as possible, and that deformation is as rigid as possible.
169 | \end{itemize}
170 | 
171 | Additionally, the $\lambda$ parameter can be adjusted to determine the relative importance of each term to the error function as a whole.
172 | 
173 | $E$ is minimised using Gauss-Newton non-linear optimisation.
174 | 
175 | \item \textbf{Surface fusion}: Now that the warp function has been obtained, applying its inverse to the live frame will warp it into the canonical model. This means they occupy the same space relative to each other, so can be directly compared.
176 | 
177 | \item \textbf{Updating the canonical model}: Now that the live frame has been warped into the canonical model, points can be added or removed from the canonical model as appropriate. 
178 | \end{enumerate}
179 | 
180 | \subsubsection{Limitations}
181 | 
182 | DynamicFusion is quite limited by its handling of topology changes; the method is robust to a change from open to closed topology, but if the surface changes from a closed topology to an open one too quickly the system will fail to reconstruct the surface.
183 | 
184 | It is also limited by unusual changes in motion. For example, it will struggle with large inter-frame motion, or the motion of occluded parts of the surface.
185 | 
186 | 
187 | \subsection{VolumeDeform}
188 | 
189 | VolumeDeform\cite{volumedeform} is different to the other examined techniques in that it doesn't rely solely on depth information. It also uses the RGB data probided by the sensor to extract additional information about the live frame.\\ 
190 | 
191 | 
192 | \subsubsection{Principles of operation}
193 | 
194 | VolumeDeform operates similarly to the other methods outlined, in that it involves a loop that repeatedly acquires new knowledge and integrates it with an existing model.
195 | 
196 | \begin{enumerate}
197 | \item \textbf{Mesh extraction}: The signed distance field representing the system's model of the scene is used to generate a deformed 3D mesh using Marching Cubes
198 | \item \textbf{Correspondence detection}: The mesh is rendered to produce a depth map which can be used to generate correspondences. 
199 | 
200 | Additionally, SIFT features are detected in the live frame and stored by the system. Since SIFT features are robust to changes in transformation, they can be recognised again in later frames. This provides extra information to inform the process of correspondence detection. 
201 | 
202 | \item \textbf{Deformation optimisation}: The deformation field is optimised so that it is consistent with the observed values of colour and depth.
203 | \end{enumerate}
204 | 
205 | \subsubsection{Limitations}
206 | 
207 | VolumeDeform comes with several limitations. Firstly, the feature tracking provided by the SIFT feature descriptor is not adequate when significant transformations are applied. SIFT feature descriptors are robust, but not robust enough given the computational resources available.
208 | 
209 | Additionally, high levels of deformation are not handled well by the system, due to over-distribution of deformation. This is as a result of the regularizer used by the system, which does not account for local rigidity.
210 | 
211 | 
212 | \subsection{KillingFusion}
213 | 
214 | KillingFusion\cite{killingfusion} operates very similarly to DynamicFusion, but introduces a much more sophisticated error function for estimating the warp field.
215 | 
216 | \begin{align*}
217 | E_{non-rigid}(\Psi) &= E_{data}(\Psi) + \omega_kE_{Killing}(\Psi) + \omega_sE_{level-set}(\Psi)\\
218 | E'_{data}(\Psi) &= (\phi_n(\Psi) - \phi_{global}) \nabla_{\phi_n}(\Psi)\\
219 | E'_{Killing}(\Psi) &= 2H_{uvw}(vec(J^{\top}_{\Psi}) vec\big(J_{\Psi})\big)\begin{pmatrix}1\\\gamma \end{pmatrix}\\
220 | E'_{level-set}(\Psi) &= \frac{|\nabla_{\phi_n}(\Psi)| - 1}{|\nabla_{\phi_n}(\Psi)|_\epsilon}H_{\phi_n}(\Psi)\nabla_{\phi_n}(\Psi)
221 | \end{align*}
222 | 
223 | This error function imposes three requirements on a potential solution:
224 | \begin{itemize}
225 | \item \textbf{Data term}: Similar to DynamicFusion, contains a component for non-rigid transformation. 
226 | \item \textbf{Killing condition}: This term requires the flow field to be a Killing vector field. If it satisifies this condition, then it will produce isometric motion Since a perfect isometric motion would impose only rigid transformations, this term only needs to be minimised so that the motion is approximately rigid. Additionally, the requirement has been weakened to make it computationally more efficient.
227 | \item \textbf{Level set property}: The minimisation of this term ensures that the gradient of the SDF function remains approximately 1
228 | \end{itemize}
229 | 
230 | The additional parameters $\omega_k$ and $\omega_s$ are used to adjust the relative importance of the three terms when calculating $E_{non-rigid}$.
231 | 
232 | This error function is also much more easily parallelised, leading to greatly increased performance.\\
233 | 
234 | 
235 | \subsubsection{Limitations}
236 | 
237 | KillingFusion has been shown to handle fast inter-frame motion and topological changes much better than alternatives such as DynamicFusion or VolumeDeform. This means that it is more desirable as a solution than the other methods outlined here.
238 | 
239 | It does however still suffer from the same limitations as all of these methods in that it can only be applied to a small volume until further improvements to the method are made.
240 | 
241 | \newpage
242 | \section{Project Plan}
243 | 
244 | \subsection{Milestones}
245 | \begin{itemize}
246 | \item February: Fully understand codebase, start programming.
247 | \item March: Reliably estimate warp field.
248 | \item April: Produce correct behaviour for DynamicFusion, not necessarily real-time.
249 | \item May: Have code running on GPU using CUDA, produce real-time solution.
250 | \item 11th May: Start working on final report.
251 | \item June: All coding finished.
252 | \item 18th June: Final report submission.
253 | \end{itemize}
254 | 
255 | \subsection{Minimum Viable Product}
256 | 
257 | In a worst-case scenario, I may fail to meet the majority of the targets I set myself in terms of implementation as well as performance. In this case, I would be able to fall back on an offline implementation of DynamicFusion. This would mean I could demonstrate the functionality of the system, without needing it to operate in real-time.
258 | 
259 | \subsection{Possible Extensions}
260 | 
261 | As a possible extension to the project, a tool to visualise the canonical model could be produced. In particular, to see how the canonical model evolves in real-time would be very useful to the user.
262 | 
263 | 
264 | \section{Evaluation Plan}
265 | \subsection{Required Functionality}
266 | 
267 | At a bare minimum, the system must be able to reconstruct a dynamic scene, not necessarily in real-time. 
268 | 
269 | Once the correct behaviour has been implemented, then the system can be optimised to work in real-time. This is the scenario I would expect to operate in.
270 | 
271 | If the project progresses well, the system could be tested under unfavourable conditions, such as large inter-frame movement or rapid topology changes. It could then be improved to handle these cases more accurately, perhaps incorporating techniques from algorithms such as KillingFusion.
272 | 
273 | If the project has progressed excellently, several extensions can be pursued as outlined in the last section.
274 | 
275 | 
276 | \subsection{Performance Metrics}
277 | 
278 | It is very simple to evaluate the performance of the final system, as its efficiency is clearly observable in the average framerate it achieves.
279 | 
280 | Ideally, the system will be able to run live and smoothly. This would mean achieving the same framerate as the Kinect sensor, which is 30Hz.  Practically speaking, it is likely that the system will not achieve such high performance. Therefore, a reduced target framerate of approximately 10Hz or over can be considered more achievable, while still providing a live ....
281 | 
282 | In the case that this target is not reached either, I will aim to produce an offline solution as a fall-back. Although the processing time will cease to be of such high importance in this scenario, it can still be used as a valid metric of the system's efficiency.
283 | 
284 | 
285 | \bibliographystyle{plain}
286 | \bibliography{report}
287 | 
288 | \end{document}
289 | 


--------------------------------------------------------------------------------
/src/deform_node.cpp:
--------------------------------------------------------------------------------
 1 | #include "deform_node.h"
 2 | 
 3 | #include "warp_field.h"
 4 | 
 5 | #include <algorithm>
 6 | 
 7 | deform_node_t::deform_node_t(warp_field_t * wf, point_t p, dual_quat_t dq, float w){
 8 |     warp_field = wf;
 9 |     position = p;
10 |     transform = dq;
11 |     radial_weight = w;
12 | }
13 | 
14 | float
15 | deform_node_t::regularisation(float phi){
16 |     // TODO: this knn search will return this node as well.
17 |     //       the terms cancel so it will end up being zero, but may want to
18 |     //       remove for efficiency or check an extra one to compensate.
19 |     auto neighbours = warp_field->find_neighbours(position);
20 |     
21 |     float result = 0.0f;
22 |     for (deform_node_t * neighbour : neighbours){
23 |         float alpha = std::max(radial_weight, neighbour->radial_weight);
24 |     }
25 | 
26 |     return result;
27 | }
28 | 
29 | float
30 | deform_node_t::weight(point_t p){
31 |     //TODO
32 |     return 0;
33 | }
34 | 
35 | dual_quat_t 
36 | deform_node_t::get_transform(){
37 |     return transform;
38 | }
39 | 
40 | point_t 
41 | deform_node_t::get_position(){
42 |     return position;
43 | }
44 | 
45 | 
46 | 


--------------------------------------------------------------------------------
/src/dual_quat.cpp:
--------------------------------------------------------------------------------
 1 | #include "dual_quat.h"
 2 | 
 3 | #include <cmath>
 4 | 
 5 | dual_quat_t::dual_quat_t(){
 6 | 
 7 | }
 8 | 
 9 | dual_quat_t::dual_quat_t(quat_t real, quat_t dual){
10 |     this->real = real;
11 |     this->dual = dual;
12 | }
13 | 
14 | dual_quat_t
15 | dual_quat_t::operator*(float scale){
16 |     return dual_quat_t(real, dual * scale);
17 | }
18 | 
19 | void
20 | dual_quat_t::operator+=(const dual_quat_t & other){
21 |     real += other.real;
22 |     dual += other.dual;
23 | }
24 | 
25 | void
26 | dual_quat_t::normalise(){
27 |     float dot = 
28 |         real.R_component_1() * real.R_component_1() +
29 |         real.R_component_2() * real.R_component_2() +
30 |         real.R_component_3() * real.R_component_3() +
31 |         real.R_component_4() * real.R_component_4();
32 |     real /= sqrt(dot);
33 | }
34 | 


--------------------------------------------------------------------------------
/src/dynamic_fusion.cpp:
--------------------------------------------------------------------------------
 1 | #include "dynamic_fusion.h"
 2 | 
 3 | #include <ceres/ceres.h>
 4 | #include <iostream>
 5 | 
 6 | using namespace ceres;
 7 | 
 8 | struct CostFunctor {
 9 |   template <typename T> bool operator()(const T* const x, T* residual) const {
10 |     residual[0] = 10.0 - x[0];
11 |     return true;
12 |   }
13 | };
14 | 
15 | dynamic_fusion_t::dynamic_fusion_t(){
16 | 
17 | }
18 | 
19 | dynamic_fusion_t::~dynamic_fusion_t(){
20 | 
21 | }
22 | 
23 | void
24 | dynamic_fusion_t::execute(){
25 |   // The variable to solve for with its initial value. It will be
26 |   // mutated in place by the solver.
27 |   double x = 0.5;
28 |   const double initial_x = x;
29 | 
30 |   // Build the problem.
31 |   Problem problem;
32 | 
33 |   // Set up the only cost function (also known as residual). This uses
34 |   // auto-differentiation to obtain the derivative (jacobian).
35 |   CostFunction* cost_function =
36 |       new AutoDiffCostFunction<CostFunctor, 1, 1>(new CostFunctor);
37 |   problem.AddResidualBlock(cost_function, NULL, &x);
38 | 
39 |   // Run the solver!
40 |   Solver::Options options;
41 |   options.minimizer_progress_to_stdout = true;
42 |   Solver::Summary summary;
43 |   Solve(options, &problem, &summary);
44 | 
45 |   std::cout << summary.BriefReport() << "\n";
46 |   std::cout << "x : " << initial_x
47 |             << " -> " << x << "\n";
48 | 
49 | }
50 | 


--------------------------------------------------------------------------------
/src/main.cpp:
--------------------------------------------------------------------------------
 1 | #include "dynamic_fusion.h"
 2 | 
 3 | #include <ceres/ceres.h>
 4 | 
 5 | int main(){
 6 |     dynamic_fusion_t d_fusion;
 7 |     d_fusion.execute();
 8 |     return 0;
 9 | }
10 | 


--------------------------------------------------------------------------------
/src/point.cpp:
--------------------------------------------------------------------------------
 1 | #include "point.h"
 2 | 
 3 | point_t::point_t() : point_t(0, 0, 0){
 4 | 
 5 | }
 6 | 
 7 | point_t::point_t(float x, float y, float z){
 8 |     elems[0] = x;
 9 |     elems[1] = y;
10 |     elems[2] = z;
11 | }
12 | 
13 | void
14 | point_t::unpack(float * array){
15 |     if (array != nullptr){
16 |         for (int i = 0; i < 3; i++){
17 |             array[i] = elems[i];
18 |         }
19 |     }
20 | }
21 | 
22 | float 
23 | point_t::get_element(int index){
24 |     if (index < 0 || index > 3){
25 |         // AAAAH
26 |         return 0;
27 |     } else {
28 |         return elems[index];
29 |     }
30 | }
31 | 
32 | std::string
33 | point_t::to_string(){
34 |     return "point(" + std::to_string(elems[0]) + ", " + std::to_string(elems[1]) + ", " + std::to_string(elems[2])  + ")";
35 | }
36 | 


--------------------------------------------------------------------------------
/src/saxpy.cu:
--------------------------------------------------------------------------------
 1 | #include <stdio.h>
 2 | 
 3 | __global__
 4 | void saxpy(int n, float a, float *x, float *y)
 5 | {
 6 |   int i = blockIdx.x*blockDim.x + threadIdx.x;
 7 |   if (i < n) y[i] = a*x[i] + y[i];
 8 | }
 9 | 
10 | int main(void)
11 | {
12 |   int N = 1<<20;
13 |   float *x, *y, *d_x, *d_y;
14 |   x = (float*)malloc(N*sizeof(float));
15 |   y = (float*)malloc(N*sizeof(float));
16 | 
17 |   cudaMalloc(&d_x, N*sizeof(float)); 
18 |   cudaMalloc(&d_y, N*sizeof(float));
19 | 
20 |   for (int i = 0; i < N; i++) {
21 |     x[i] = 1.0f;
22 |     y[i] = 2.0f;
23 |   }
24 | 
25 |   cudaMemcpy(d_x, x, N*sizeof(float), cudaMemcpyHostToDevice);
26 |   cudaMemcpy(d_y, y, N*sizeof(float), cudaMemcpyHostToDevice);
27 | 
28 |   // Perform SAXPY on 1M elements
29 |   saxpy<<<(N+255)/256, 256>>>(N, 2.0f, d_x, d_y);
30 | 
31 |   cudaMemcpy(y, d_y, N*sizeof(float), cudaMemcpyDeviceToHost);
32 | 
33 |   float maxError = 0.0f;
34 |   for (int i = 0; i < N; i++)
35 |     maxError = max(maxError, abs(y[i]-4.0f));
36 |   printf("Max error: %f\n", maxError);
37 | 
38 |   cudaFree(d_x);
39 |   cudaFree(d_y);
40 |   free(x);
41 |   free(y);
42 | }
43 | 


--------------------------------------------------------------------------------
/src/warp_field.cpp:
--------------------------------------------------------------------------------
 1 | #include "warp_field.h"
 2 | 
 3 | #include <iostream>
 4 | 
 5 | warp_field_t::warp_field_t(){
 6 |     kd_tree = new kd_tree_t(3, *this, KDTreeSingleIndexAdaptorParams(10)); 
 7 | }
 8 | 
 9 | warp_field_t::~warp_field_t(){
10 |     delete kd_tree;
11 | }
12 | 
13 | std::vector<deform_node_t *>
14 | warp_field_t::find_neighbours(point_t p){
15 |     // output parameters from knnSearch, representing the indices of the closest points
16 |     // and the distance to those points, which for now are discarded
17 |     size_t neighbour_indices[NUM_NEIGHBOURS];
18 |     float neighbour_distances[NUM_NEIGHBOURS];    
19 | 
20 |     // unpack vector into float array
21 |     float query[3];
22 |     p.unpack(query);
23 | 
24 |     // query KD-tree
25 |     
26 |     // TODO: this should be after every modification of the node vector
27 |     kd_tree->buildIndex();
28 |     int number_found = kd_tree->knnSearch(query, NUM_NEIGHBOURS, neighbour_indices, neighbour_distances);
29 |     
30 |     // create map to return
31 |     std::vector<deform_node_t *> result;
32 |     for (int i = 0; i < number_found; i++){
33 |         result.push_back(nodes[neighbour_indices[i]]);
34 |     }
35 | 
36 |     return result;
37 | }
38 | 
39 | dual_quat_t 
40 | warp_field_t::dual_quaternion_blending(point_t p){
41 |     // find the N closest neighbours to the given point
42 |     auto neighbours = find_neighbours(p);
43 |  
44 |     // sum over all neighbours transforms adjusted by their weight
45 |     dual_quat_t q;
46 |     for (auto neighbour : neighbours){
47 |         // NOTE: dynfu implementation has this as a product (*=)
48 |         //       but dynamicfusion paper specifies a summation?
49 |         q += neighbour->get_transform() * neighbour->weight(p);
50 |     }
51 | 
52 |     // normalize and return
53 |     q.normalise();
54 |     return q;    
55 | }
56 | 
57 | size_t
58 | warp_field_t::kdtree_get_point_count() const {
59 |     return nodes.size();
60 | }
61 | 
62 | float 
63 | warp_field_t::kdtree_get_pt(const size_t idx, int dim) const {
64 |     return nodes[idx]->get_position().get_element(dim);
65 | }
66 | 
67 | void
68 | warp_field_t::update_warp_field(){
69 |     kd_tree->buildIndex();
70 | }
71 | 


--------------------------------------------------------------------------------