├── .gitignore ├── .gitmodules ├── LICENSE ├── README.md ├── configs └── run_odometry.yaml ├── data └── .gitkeep ├── libs ├── kinetics.py └── misc.py ├── models ├── __init__.py ├── early_fusion.py ├── late_fusion.py ├── middle_fusion.py ├── multi_layer_fusion.py └── util.py ├── requirements.txt └── run_odometry.py /.gitignore: -------------------------------------------------------------------------------- 1 | __pycache__ 2 | -------------------------------------------------------------------------------- /.gitmodules: -------------------------------------------------------------------------------- 1 | [submodule "third_party/monodepth2"] 2 | path = third_party/monodepth2 3 | url = https://github.com/nianticlabs/monodepth2 4 | [submodule "third_party/CEN"] 5 | path = third_party/CEN 6 | url = https://github.com/yikaiw/CEN 7 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # MLF-VO 2 | 3 | This repo implements the network described in the ICRA2022 paper: 4 | 5 | [Self-Supervised Ego-Motion Estimation Based on Multi-Layer Fusion of RGB and Inferred Depth](https://arxiv.org/pdf/2203.01557.pdf) 6 | 7 | Zijie Jiang, Hajime Taira, Naoyuki Miyashita and Masatoshi Okutomi 8 | 9 | Please also visit our [project page](http://www.ok.sc.e.titech.ac.jp/res/MLF-VO/). 10 | 11 | If you find our work useful for your research, please consider citing the following paper: 12 | 13 | ``` 14 | @inproceedings{jiang2022mlfvo, 15 | title={Self-Supervised Ego-Motion Estimation Based on Multi-Layer Fusion of RGB and Inferred Depth}, 16 | author={Jiang, Zijie and Taira, Hajime and Miyashita, Naoyuki and Okutomi, Masatoshi}, 17 | booktitle={2022 IEEE International Conference on Robotics and Automation (ICRA)}, 18 | year={2022} 19 | } 20 | ``` 21 | 22 | # Updates 23 | 24 | - [x] 2022.05.20. Pretrained weights and demo for testing are open. 25 | 26 | # Requirements 27 | 28 | Our experiments are conducted on a machine installed with Ubuntu 18.04, Pytorch 1.7.1, CUDA 10.1. You can install other required packages by: 29 | 30 | ``` bash 31 | pip install requirements.txt 32 | ``` 33 | 34 | Since our demo partly depends on [Monodepth2](https://github.com/nianticlabs/monodepth2) and [CEN](https://github.com/yikaiw/CEN), clone this repository with `--recurse` option. Please also refer their original repositories. 35 | 36 | # Prepare dataset 37 | 38 | Please download the KITTI odometry benchmark from their [site](http://www.cvlibs.net/datasets/kitti/eval_odometry.php) and create a soft link by executing the following command: 39 | 40 | ``` bash 41 | ln -s /path/to/dataset/data_odometry_color/dataset data 42 | ``` 43 | 44 | # Odometry evaluation 45 | 46 | Download and place the pretrained weights of models following: 47 | 48 | ``` bash 49 | wget http://www.ok.sc.e.titech.ac.jp/res/MLF-VO/models.zip 50 | unzip -d data models.zip 51 | rm models.zip 52 | ``` 53 | 54 | The demo specifies several paths and configurations in `configs/run_odometry.yaml`. Please modify line 2, 6, or 9 to use your own paths to dataset, model weights, or outputs. 55 | 56 | ``` bash 57 | python run_odometry -c configs/run_odometry.yaml 58 | ``` 59 | 60 | The estimated trajectory will be exported to `data/outputs/09.txt`. Each line in the file represents the estimated camera, corresponding to each frame in the sequence, in the following format: 61 | 62 | ``` 63 | T11 T12 T13 T14 T21 T22 T23 T24 T31 T32 T33 T34 64 | ... 65 | ``` 66 | 67 | where T11, T12, ... , T34 are the elements of 3x4 camera transformation matrix (from camera to world coordinate system): 68 | 69 | ``` 70 | T11 T12 T13 T14 71 | T21 T22 T23 T24 72 | T31 T32 T33 T34 73 | ``` 74 | 75 | We recommend to use this [toolbox](https://github.com/Huangying-Zhan/kitti-odom-eval) to evaluate the inferred trajectory. 76 | 77 | # Training 78 | 79 | Please checkout to the `dev` branch, and run 80 | 81 | ``` 82 | python train.py --data_path ~/Documents/datasets/kitti_odom/dataset --log_dir ~/Documents/MLF-VO/KITTI --model_name test --num_epochs 20 --dataset kitti_odom --split odom 83 | ``` 84 | 85 | Please change L55 in `trainer.py` to the wanted type of the PoseNet. 86 | 87 | ```python 88 | self.models['pose'] = posenet_type_dict[config_dict.model.type]() 89 | ``` 90 | 91 | # Acknowledgement 92 | 93 | We are grateful to the authors of [Monodepth2](https://github.com/nianticlabs/monodepth2) and [CEN](https://github.com/yikaiw/CEN) for publicly sharing their codes. 94 | 95 | # License 96 | 97 | MLF-VO is released under GPLv3 License. 98 | -------------------------------------------------------------------------------- /configs/run_odometry.yaml: -------------------------------------------------------------------------------- 1 | data: 2 | path: 'data/dataset/sequences/09' # path to dataset for evaluation 3 | height: 192 4 | width: 640 5 | model: 6 | path: 'data/models/RGB+Depth_multi_layer' # path to models for evaluation 7 | type: 'multi_layer' 8 | output: 9 | path: 'data/outputs' # path for saving outputs 10 | save_depth: False 11 | -------------------------------------------------------------------------------- /data/.gitkeep: -------------------------------------------------------------------------------- 1 | data -------------------------------------------------------------------------------- /libs/kinetics.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | 4 | def rot_from_euler_angle(angle): 5 | """Convert euler angles to rotation matrix. 6 | Reference: https://github.com/pulkitag/pycaffe-utils/blob/master/rot_utils.py#L174 7 | Args: 8 | angle: rotation angle along 3 axis (in radians) -- size = [B, 3] 9 | Returns: 10 | Rotation matrix corresponding to the euler angles -- size = [B, 3, 3] 11 | """ 12 | B = angle.size(0) 13 | x, y, z = angle[:, 0], angle[:, 1], angle[:, 2] 14 | 15 | cosz = torch.cos(z) 16 | sinz = torch.sin(z) 17 | 18 | zeros = torch.zeros_like(cosz) 19 | ones = torch.ones_like(cosz) 20 | zmat = torch.stack([cosz, -sinz, zeros, 21 | sinz, cosz, zeros, 22 | zeros, zeros, ones], dim=1).reshape(B, 3, 3) 23 | 24 | cosy = torch.cos(y) 25 | siny = torch.sin(y) 26 | 27 | ymat = torch.stack([cosy, zeros, siny, 28 | zeros, ones, zeros, 29 | -siny, zeros, cosy], dim=1).reshape(B, 3, 3) 30 | 31 | cosx = torch.cos(x) 32 | sinx = torch.sin(x) 33 | 34 | xmat = torch.stack([ones, zeros, zeros, 35 | zeros, cosx, -sinx, 36 | zeros, sinx, cosx], dim=1).reshape(B, 3, 3) 37 | 38 | rotMat = xmat @ ymat @ zmat 39 | 40 | rot = torch.zeros((angle.shape[0], 4, 4)).to(device=angle.device) 41 | rot[:, :3, :3] = rotMat 42 | rot[:, 3, 3] = 1 43 | 44 | return rot 45 | 46 | 47 | def from_euler_t(euler_angle, translation, invert=False): 48 | T = rot_from_euler_angle(euler_angle) 49 | T[:, :3, 3] = translation 50 | 51 | if invert: 52 | T = T.pinverse() 53 | 54 | return T 55 | -------------------------------------------------------------------------------- /libs/misc.py: -------------------------------------------------------------------------------- 1 | import os 2 | 3 | 4 | def create_dir_if_not_exist(dir): 5 | if not os.path.exists(dir): 6 | os.makedirs 7 | -------------------------------------------------------------------------------- /models/__init__.py: -------------------------------------------------------------------------------- 1 | from .early_fusion import create_posenet_early 2 | from .late_fusion import create_posenet_late 3 | from .middle_fusion import create_posenet_middle 4 | from .multi_layer_fusion import create_posenet_multi_layer 5 | -------------------------------------------------------------------------------- /models/early_fusion.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | from torch.nn.modules.batchnorm import BatchNorm2d 4 | 5 | 6 | def conv3x3(in_planes, out_planes, stride=1, bias=False): 7 | "3x3 convolution with padding" 8 | return nn.Conv2d(in_planes, out_planes, kernel_size=3, 9 | stride=stride, padding=1, bias=bias) 10 | 11 | 12 | def conv1x1(in_planes, out_planes, stride=1, bias=False): 13 | "1x1 convolution" 14 | return nn.Conv2d(in_planes, out_planes, kernel_size=1, 15 | stride=stride, padding=0, bias=bias) 16 | 17 | 18 | class BasicBlock(nn.Module): 19 | expansion = 1 20 | def __init__(self, inplanes, planes, stride=1, downsample=None): 21 | super(BasicBlock, self).__init__() 22 | self.conv1 = conv3x3(inplanes, planes, stride) 23 | self.bn1 = BatchNorm2d(planes) 24 | self.relu = nn.ReLU(inplace=True) 25 | self.conv2 = conv3x3(planes, planes) 26 | self.bn2 = BatchNorm2d(planes) 27 | 28 | self.stride = stride 29 | self.downsample = downsample 30 | 31 | def forward(self, x): 32 | residual = x 33 | 34 | out = self.conv1(x) 35 | out = self.bn1(out) 36 | out = self.relu(out) 37 | 38 | out = self.conv2(out) 39 | out = self.bn2(out) 40 | 41 | if self.downsample is not None: 42 | residual = self.downsample(x) 43 | 44 | out = out + residual 45 | out = self.relu(out) 46 | 47 | return out 48 | 49 | 50 | class PoseNet(nn.Module): 51 | def __init__(self, block, layers): 52 | super(PoseNet, self).__init__() 53 | 54 | self.inplanes = 64 55 | 56 | self.conv1 = nn.Conv2d(12, 64, kernel_size=7, stride=2, padding=3, bias=False) 57 | 58 | self.bn1 = nn.BatchNorm2d(64) 59 | self.relu = nn.ReLU(inplace=True) 60 | self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) 61 | self.layer1 = self._make_layer(block, 64, layers[0]) 62 | self.layer2 = self._make_layer(block, 128, layers[1], stride=2) 63 | self.layer3 = self._make_layer(block, 256, layers[2], stride=2) 64 | self.layer4 = self._make_layer(block, 512, layers[3], stride=2) 65 | 66 | self.fc1 = nn.Linear(512 * 6 * 20, 512) 67 | self.fc2 = nn.Linear(512, 512) 68 | self.fc3_t = nn.Linear(512, 3) 69 | self.fc3_r = nn.Linear(512, 3) 70 | 71 | for m in self.modules(): 72 | if isinstance(m, nn.Conv2d): 73 | nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') 74 | elif isinstance(m, nn.BatchNorm2d): 75 | nn.init.constant_(m.weight, 1) 76 | nn.init.constant_(m.bias, 0) 77 | 78 | def _make_layer(self, block, planes, num_blocks, stride=1): 79 | downsample = None 80 | if stride != 1 or self.inplanes != planes * block.expansion: 81 | downsample = nn.Sequential( 82 | conv1x1(self.inplanes, planes * block.expansion, stride=stride), 83 | BatchNorm2d(planes * block.expansion) 84 | ) 85 | 86 | layers = [] 87 | layers.append(block(self.inplanes, planes, stride, downsample)) 88 | self.inplanes = planes * block.expansion 89 | for _ in range(1, num_blocks): 90 | layers.append(block(self.inplanes, planes)) 91 | 92 | return nn.Sequential(*layers) 93 | 94 | def forward(self, x): 95 | x = torch.cat(x, dim=1) 96 | 97 | x = self.conv1(x) 98 | x = self.bn1(x) 99 | x = self.relu(x) 100 | x = self.maxpool(x) 101 | 102 | x = self.layer1(x) 103 | x = self.layer2(x) 104 | x = self.layer3(x) 105 | x = self.layer4(x) 106 | 107 | x = x.flatten(start_dim=1) 108 | 109 | x = self.fc1(x) 110 | x = self.relu(x) 111 | x = self.fc2(x) 112 | x = self.relu(x) 113 | 114 | x_t = self.fc3_t(x) 115 | x_r = self.fc3_r(x) 116 | x = torch.cat([x_r, x_t], dim=1) 117 | 118 | axisangle = x[:, 0:3] 119 | translation = x[:, 3:6] 120 | axisangle = 0.001 * axisangle.view(-1, 3) 121 | translation = 0.001 * translation.view(-1, 3) 122 | 123 | return axisangle, translation 124 | 125 | 126 | def create_posenet_early(): 127 | model = PoseNet(BasicBlock, [2, 2, 2, 2]) 128 | 129 | return model 130 | -------------------------------------------------------------------------------- /models/late_fusion.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | 5 | from .util import conv7x7Parallel, conv3x3Parallel, conv1x1Parallel, linearParallel 6 | from third_party.CEN.semantic_segmentation.models.modules import BatchNorm2dParallel, ModuleParallel 7 | 8 | 9 | class BasicBlock(nn.Module): 10 | expansion = 1 11 | def __init__(self, inplanes, planes, num_parallel, stride=1, downsample=None): 12 | super(BasicBlock, self).__init__() 13 | self.conv1 = conv3x3Parallel(inplanes, planes, num_parallel, stride) 14 | self.bn1 = BatchNorm2dParallel(planes, num_parallel) 15 | self.relu = ModuleParallel(nn.ReLU(inplace=True)) 16 | self.conv2 = conv3x3Parallel(planes, planes, num_parallel) 17 | self.bn2 = BatchNorm2dParallel(planes, num_parallel) 18 | 19 | self.num_parallel = num_parallel 20 | self.stride = stride 21 | self.downsample = downsample 22 | 23 | def forward(self, x): 24 | residual = x 25 | 26 | out = self.conv1(x) 27 | out = self.bn1(out) 28 | out = self.relu(out) 29 | 30 | out = self.conv2(out) 31 | out = self.bn2(out) 32 | 33 | if self.downsample is not None: 34 | residual = self.downsample(x) 35 | 36 | out = [out[l] + residual[l] for l in range(self.num_parallel)] 37 | out = self.relu(out) 38 | 39 | return out 40 | 41 | 42 | class PoseNet(nn.Module): 43 | def __init__(self, block, layers, num_parallel): 44 | super(PoseNet, self).__init__() 45 | 46 | self.inplanes = 64 47 | self.num_parallel = num_parallel 48 | 49 | self.conv1 = conv7x7Parallel(6, 64, 2, stride=2) 50 | 51 | self.bn1 = BatchNorm2dParallel(64, num_parallel) 52 | self.relu = ModuleParallel(nn.ReLU(inplace=True)) 53 | self.maxpool = ModuleParallel(nn.MaxPool2d(kernel_size=3, stride=2, padding=1)) 54 | self.layer1 = self._make_layer(block, 64, layers[0]) 55 | self.layer2 = self._make_layer(block, 128, layers[1], stride=2) 56 | self.layer3 = self._make_layer(block, 256, layers[2], stride=2) 57 | self.layer4 = self._make_layer(block, 512, layers[3], stride=2) 58 | 59 | self.fc1 = linearParallel(512 * 6 * 20, 512, 2) 60 | self.fc2 = linearParallel(512, 512, 2) 61 | self.fc3_t = linearParallel(512, 3, 2) 62 | self.fc3_r = linearParallel(512, 3, 2) 63 | 64 | self.alpha = nn.Parameter(torch.ones(num_parallel, requires_grad=True)) 65 | self.register_parameter('alpha', self.alpha) 66 | 67 | for m in self.modules(): 68 | if isinstance(m, nn.Conv2d): 69 | nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') 70 | elif isinstance(m, nn.BatchNorm2d): 71 | nn.init.constant_(m.weight, 1) 72 | nn.init.constant_(m.bias, 0) 73 | 74 | def _make_layer(self, block, planes, num_blocks, stride=1): 75 | downsample = None 76 | if stride != 1 or self.inplanes != planes * block.expansion: 77 | downsample = nn.Sequential( 78 | conv1x1Parallel(self.inplanes, planes * block.expansion, 2, stride=stride), 79 | BatchNorm2dParallel(planes * block.expansion, self.num_parallel) 80 | ) 81 | 82 | layers = [] 83 | layers.append(block(self.inplanes, planes, self.num_parallel, stride, downsample)) 84 | self.inplanes = planes * block.expansion 85 | for _ in range(1, num_blocks): 86 | layers.append(block(self.inplanes, planes, self.num_parallel)) 87 | 88 | return nn.Sequential(*layers) 89 | 90 | def forward(self, x): 91 | x = self.conv1(x) 92 | x = self.bn1(x) 93 | x = self.relu(x) 94 | x = self.maxpool(x) 95 | 96 | x = self.layer1(x) 97 | x = self.layer2(x) 98 | x = self.layer3(x) 99 | x = self.layer4(x) 100 | 101 | x = [_x.flatten(start_dim=1) for _x in x] 102 | 103 | x = self.fc1(x) 104 | x = self.relu(x) 105 | x = self.fc2(x) 106 | x = self.relu(x) 107 | 108 | x_t = self.fc3_t(x) 109 | x_r = self.fc3_r(x) 110 | x = [torch.cat([_x_r, _x_t], dim=1) for _x_r, _x_t in zip(x_r, x_t)] 111 | 112 | alpha_soft = F.softmax(self.alpha) 113 | axisangle = alpha_soft[0] * x[0][:, 0:3] + alpha_soft[1] * x[1][:, 0:3] 114 | translation = alpha_soft[0] * x[0][:, 3:6] + alpha_soft[1] * x[1][:, 3:6] 115 | axisangle = 0.001 * axisangle.view(-1, 3) 116 | translation = 0.001 * translation.view(-1, 3) 117 | 118 | return axisangle, translation 119 | 120 | 121 | def create_posenet_late(num_parallel=2): 122 | model = PoseNet(BasicBlock, [2, 2, 2, 2], num_parallel) 123 | 124 | return model 125 | -------------------------------------------------------------------------------- /models/middle_fusion.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | 4 | from .util import conv7x7Parallel, conv3x3Parallel, conv1x1Parallel, linearParallel 5 | from third_party.CEN.semantic_segmentation.models.modules import BatchNorm2dParallel, ModuleParallel 6 | 7 | 8 | class BasicBlock(nn.Module): 9 | expansion = 1 10 | def __init__(self, inplanes, planes, num_parallel, stride=1, downsample=None): 11 | super(BasicBlock, self).__init__() 12 | self.conv1 = conv3x3Parallel(inplanes, planes, num_parallel, stride) 13 | self.bn1 = BatchNorm2dParallel(planes, num_parallel) 14 | self.relu = ModuleParallel(nn.ReLU(inplace=True)) 15 | self.conv2 = conv3x3Parallel(planes, planes, num_parallel) 16 | self.bn2 = BatchNorm2dParallel(planes, num_parallel) 17 | 18 | self.num_parallel = num_parallel 19 | self.stride = stride 20 | self.downsample = downsample 21 | 22 | def forward(self, x): 23 | residual = x 24 | 25 | out = self.conv1(x) 26 | out = self.bn1(out) 27 | out = self.relu(out) 28 | 29 | out = self.conv2(out) 30 | out = self.bn2(out) 31 | 32 | if self.downsample is not None: 33 | residual = self.downsample(x) 34 | 35 | out = [out[l] + residual[l] for l in range(self.num_parallel)] 36 | out = self.relu(out) 37 | 38 | return out 39 | 40 | 41 | class PoseNet(nn.Module): 42 | def __init__(self, block, layers, num_parallel): 43 | super(PoseNet, self).__init__() 44 | 45 | self.inplanes = 64 46 | self.num_parallel = num_parallel 47 | 48 | self.conv1 = conv7x7Parallel(6, 64, 2, stride=2) 49 | 50 | self.bn1 = BatchNorm2dParallel(64, num_parallel) 51 | self.relu = ModuleParallel(nn.ReLU(inplace=True)) 52 | self.relu_ = nn.ReLU(inplace=True) 53 | self.maxpool = ModuleParallel(nn.MaxPool2d(kernel_size=3, stride=2, padding=1)) 54 | self.layer1 = self._make_layer(block, 64, layers[0]) 55 | self.layer2 = self._make_layer(block, 128, layers[1], stride=2) 56 | self.layer3 = self._make_layer(block, 256, layers[2], stride=2) 57 | self.layer4 = self._make_layer(block, 512, layers[3], stride=2) 58 | 59 | self.fc1 = linearParallel(512 * 6 * 20, 256, 2) 60 | self.fc2_1 = nn.Linear(512, 512) 61 | self.fc2_2 = nn.Linear(512, 512) 62 | self.soft_fusion = nn.Linear(512, 512) 63 | self.fc3_t = nn.Linear(512, 3) 64 | self.fc3_r = nn.Linear(512, 3) 65 | 66 | self.sigmoid = nn.Sigmoid() 67 | 68 | for m in self.modules(): 69 | if isinstance(m, nn.Conv2d): 70 | nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') 71 | elif isinstance(m, nn.BatchNorm2d): 72 | nn.init.constant_(m.weight, 1) 73 | nn.init.constant_(m.bias, 0) 74 | 75 | def _make_layer(self, block, planes, num_blocks, stride=1): 76 | downsample = None 77 | if stride != 1 or self.inplanes != planes * block.expansion: 78 | downsample = nn.Sequential( 79 | conv1x1Parallel(self.inplanes, planes * block.expansion, 2, stride=stride), 80 | BatchNorm2dParallel(planes * block.expansion, self.num_parallel) 81 | ) 82 | 83 | layers = [] 84 | layers.append(block(self.inplanes, planes, self.num_parallel, stride, downsample)) 85 | self.inplanes = planes * block.expansion 86 | for _ in range(1, num_blocks): 87 | layers.append(block(self.inplanes, planes, self.num_parallel)) 88 | 89 | return nn.Sequential(*layers) 90 | 91 | def forward(self, x): 92 | x = self.conv1(x) 93 | x = self.bn1(x) 94 | x = self.relu(x) 95 | x = self.maxpool(x) 96 | 97 | x = self.layer1(x) 98 | x = self.layer2(x) 99 | x = self.layer3(x) 100 | x = self.layer4(x) 101 | 102 | x = [_x.flatten(start_dim=1) for _x in x] 103 | 104 | x = self.fc1(x) 105 | x = self.relu(x) 106 | 107 | x = torch.cat(x, dim=1) 108 | mask = self.sigmoid(self.soft_fusion(x)) 109 | x = mask * x 110 | 111 | x = self.fc2_1(x) 112 | x = self.relu_(x) 113 | x = self.fc2_2(x) 114 | x = self.relu_(x) 115 | 116 | x_t = self.fc3_t(x) 117 | x_r = self.fc3_r(x) 118 | x = torch.cat([x_r, x_t], dim=1) 119 | 120 | axisangle = x[:, 0:3] 121 | translation = x[:, 3:6] 122 | axisangle = 0.01 * axisangle.view(-1, 3) 123 | translation = 0.01 * translation.view(-1, 3) 124 | 125 | return axisangle, translation 126 | 127 | 128 | def create_posenet_middle(num_parallel=2): 129 | model = PoseNet(BasicBlock, [2, 2, 2, 2], num_parallel) 130 | 131 | return model 132 | -------------------------------------------------------------------------------- /models/multi_layer_fusion.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | 5 | from .util import conv3x3, conv1x1 6 | from third_party.CEN.semantic_segmentation.models.modules import Exchange, BatchNorm2dParallel, ModuleParallel 7 | 8 | 9 | class BasicBlock(nn.Module): 10 | expansion = 1 11 | def __init__(self, inplanes, planes, num_parallel, bn_threshold, stride=1, downsample=None): 12 | super(BasicBlock, self).__init__() 13 | self.conv1 = conv3x3(inplanes, planes, stride) 14 | self.bn1 = BatchNorm2dParallel(planes, num_parallel) 15 | self.relu = ModuleParallel(nn.ReLU(inplace=True)) 16 | self.conv2 = conv3x3(planes, planes) 17 | self.bn2 = BatchNorm2dParallel(planes, num_parallel) 18 | 19 | self.num_parallel = num_parallel 20 | self.stride = stride 21 | self.downsample = downsample 22 | 23 | self.exchange = Exchange() 24 | self.bn_threshold = bn_threshold 25 | self.bn2_list = [] 26 | for module in self.bn2.modules(): 27 | if isinstance(module, nn.BatchNorm2d): 28 | self.bn2_list.append(module) 29 | 30 | def forward(self, x): 31 | residual = x 32 | 33 | out = self.conv1(x) 34 | out = self.bn1(out) 35 | out = self.relu(out) 36 | 37 | out = self.conv2(out) 38 | out = self.bn2(out) 39 | if len(x) > 1: 40 | out = self.exchange(out, self.bn2_list, self.bn_threshold) 41 | 42 | if self.downsample is not None: 43 | residual = self.downsample(x) 44 | 45 | out = [out[l] + residual[l] for l in range(self.num_parallel)] 46 | out = self.relu(out) 47 | 48 | return out 49 | 50 | 51 | class PoseNet(nn.Module): 52 | def __init__(self, block, layers, num_parallel, bn_threshold=2e-2): 53 | super(PoseNet, self).__init__() 54 | 55 | self.inplanes = 64 56 | self.num_parallel = num_parallel 57 | 58 | self.conv1 = ModuleParallel(nn.Conv2d(6, 64, kernel_size=7, stride=2, padding=3, bias=False)) 59 | 60 | self.bn1 = BatchNorm2dParallel(64, num_parallel) 61 | self.relu = ModuleParallel(nn.ReLU(inplace=True)) 62 | self.maxpool = ModuleParallel(nn.MaxPool2d(kernel_size=3, stride=2, padding=1)) 63 | self.layer1 = self._make_layer(block, 64, layers[0], bn_threshold) 64 | self.layer2 = self._make_layer(block, 128, layers[1], bn_threshold, stride=2) 65 | self.layer3 = self._make_layer(block, 256, layers[2], bn_threshold, stride=2) 66 | self.layer4 = self._make_layer(block, 512, layers[3], bn_threshold, stride=2) 67 | 68 | self.fc1 = ModuleParallel(nn.Linear(512 * 6 * 20, 512)) 69 | self.fc2 = ModuleParallel(nn.Linear(512, 512)) 70 | self.fc3_t = ModuleParallel(nn.Linear(512, 3)) 71 | self.fc3_r = ModuleParallel(nn.Linear(512, 3)) 72 | 73 | self.alpha = nn.Parameter(torch.ones(num_parallel, requires_grad=True)) 74 | self.register_parameter('alpha', self.alpha) 75 | 76 | for m in self.modules(): 77 | if isinstance(m, nn.Conv2d): 78 | nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') 79 | elif isinstance(m, nn.BatchNorm2d): 80 | nn.init.constant_(m.weight, 1) 81 | nn.init.constant_(m.bias, 0) 82 | 83 | def _make_layer(self, block, planes, num_blocks, bn_threshold, stride=1): 84 | downsample = None 85 | if stride != 1 or self.inplanes != planes * block.expansion: 86 | downsample = nn.Sequential( 87 | conv1x1(self.inplanes, planes * block.expansion, stride=stride), 88 | BatchNorm2dParallel(planes * block.expansion, self.num_parallel) 89 | ) 90 | 91 | layers = [] 92 | layers.append(block(self.inplanes, planes, self.num_parallel, bn_threshold, stride, downsample)) 93 | self.inplanes = planes * block.expansion 94 | for _ in range(1, num_blocks): 95 | layers.append(block(self.inplanes, planes, self.num_parallel, bn_threshold)) 96 | 97 | return nn.Sequential(*layers) 98 | 99 | def forward(self, x): 100 | x = self.conv1(x) 101 | x = self.bn1(x) 102 | x = self.relu(x) 103 | x = self.maxpool(x) 104 | 105 | x = self.layer1(x) 106 | x = self.layer2(x) 107 | x = self.layer3(x) 108 | x = self.layer4(x) 109 | 110 | x = [_x.flatten(start_dim=1) for _x in x] 111 | 112 | x = self.fc1(x) 113 | x = self.relu(x) 114 | x = self.fc2(x) 115 | x = self.relu(x) 116 | 117 | x_t = self.fc3_t(x) 118 | x_r = self.fc3_r(x) 119 | x = [torch.cat([_x_r, _x_t], dim=1) for _x_r, _x_t in zip(x_r, x_t)] 120 | 121 | alpha_soft = F.softmax(self.alpha, dim=0) 122 | axisangle = alpha_soft[0] * x[0][:, 0:3] + alpha_soft[1] * x[1][:, 0:3] 123 | translation = alpha_soft[0] * x[0][:, 3:6] + alpha_soft[1] * x[1][:, 3:6] 124 | axisangle = 0.001 * axisangle.view(-1, 3) 125 | translation = 0.001 * translation.view(-1, 3) 126 | 127 | return axisangle, translation 128 | 129 | 130 | def create_posenet_multi_layer(num_parallel=2, bn_threshold=2e-2): 131 | model = PoseNet(BasicBlock, [2, 2, 2, 2], num_parallel, bn_threshold) 132 | 133 | return model 134 | -------------------------------------------------------------------------------- /models/util.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | 3 | from third_party.CEN.semantic_segmentation.models.modules import ModuleParallel 4 | 5 | 6 | def conv3x3(in_planes, out_planes, stride=1, bias=False): 7 | "3x3 convolution with padding" 8 | return ModuleParallel(nn.Conv2d(in_planes, out_planes, kernel_size=3, 9 | stride=stride, padding=1, bias=bias)) 10 | 11 | 12 | def conv1x1(in_planes, out_planes, stride=1, bias=False): 13 | "1x1 convolution" 14 | return ModuleParallel(nn.Conv2d(in_planes, out_planes, kernel_size=1, 15 | stride=stride, padding=0, bias=bias)) 16 | 17 | 18 | class conv7x7Parallel(nn.Module): 19 | def __init__(self, in_planes, out_planes, num_parallel, stride=1, bias=False): 20 | super(conv7x7Parallel, self).__init__() 21 | for i in range(num_parallel): 22 | setattr(self, 'conv7x7_' + str(i), nn.Conv2d(in_planes, out_planes, kernel_size=7, 23 | stride=stride, padding=3, bias=bias)) 24 | 25 | def forward(self, x_parallel): 26 | return [getattr(self, 'conv7x7_' + str(i))(x) for i, x in enumerate(x_parallel)] 27 | 28 | 29 | class conv3x3Parallel(nn.Module): 30 | def __init__(self, in_planes, out_planes, num_parallel, stride=1, bias=False): 31 | super(conv3x3Parallel, self).__init__() 32 | for i in range(num_parallel): 33 | setattr(self, 'conv3x3_' + str(i), nn.Conv2d(in_planes, out_planes, kernel_size=3, 34 | stride=stride, padding=1, bias=bias)) 35 | 36 | def forward(self, x_parallel): 37 | return [getattr(self, 'conv3x3_' + str(i))(x) for i, x in enumerate(x_parallel)] 38 | 39 | 40 | class conv1x1Parallel(nn.Module): 41 | def __init__(self, in_planes, out_planes, num_parallel, stride=1, bias=False): 42 | super(conv1x1Parallel, self).__init__() 43 | for i in range(num_parallel): 44 | setattr(self, 'conv1x1_' + str(i), nn.Conv2d(in_planes, out_planes, kernel_size=1, 45 | stride=stride, padding=0, bias=bias)) 46 | 47 | def forward(self, x_parallel): 48 | return [getattr(self, 'conv1x1_' + str(i))(x) for i, x in enumerate(x_parallel)] 49 | 50 | 51 | class linearParallel(nn.Module): 52 | def __init__(self, in_features, out_features, num_parallel) -> None: 53 | super().__init__() 54 | for i in range(num_parallel): 55 | setattr(self, 'linear_' + str(i), nn.Linear(in_features, out_features)) 56 | 57 | def forward(self, x_parallel): 58 | return [getattr(self, 'linear_' + str(i))(x) for i, x in enumerate(x_parallel)] 59 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | easydict==1.9 2 | numpy==1.18.5 3 | PyYAML==6.0 4 | torch==1.7.1+cu101 5 | torchvision==0.8.2+cu101 6 | -------------------------------------------------------------------------------- /run_odometry.py: -------------------------------------------------------------------------------- 1 | import os 2 | import argparse 3 | from pathlib import Path 4 | import yaml 5 | from easydict import EasyDict as edict 6 | import torch 7 | from torchvision import transforms 8 | import numpy as np 9 | from PIL import Image 10 | from tqdm import tqdm 11 | 12 | from models import create_posenet_early, create_posenet_late, create_posenet_middle, create_posenet_multi_layer 13 | from libs.kinetics import from_euler_t 14 | from libs.misc import create_dir_if_not_exist 15 | from third_party.monodepth2.networks import ResnetEncoder 16 | from third_party.monodepth2.networks import DepthDecoder 17 | 18 | 19 | posenet_type_dict = { 20 | 'early': create_posenet_early, 21 | 'late': create_posenet_late, 22 | 'middle': create_posenet_middle, 23 | 'multi_layer': create_posenet_multi_layer 24 | } 25 | 26 | 27 | def pil_loader(path): 28 | # open path as file to avoid ResourceWarning 29 | # (https://github.com/python-pillow/Pillow/issues/835) 30 | with open(path, 'rb') as f: 31 | with Image.open(f) as img: 32 | return img.convert('RGB') 33 | 34 | 35 | class KIITIOdomDataset: 36 | def __init__(self, config_dict) -> None: 37 | super().__init__() 38 | 39 | self.seq_dir = os.path.join(config_dict.data.path, 'image_2') 40 | self.height = config_dict.data.height 41 | self.width = config_dict.data.width 42 | 43 | self.filename_list = sorted(os.listdir(self.seq_dir)) 44 | self.loader = pil_loader 45 | self.resize = transforms.Resize((self.height, self.width), interpolation=Image.ANTIALIAS) 46 | self.to_tensor = transforms.ToTensor() 47 | 48 | def __len__(self): 49 | return len(self.filename_list) 50 | 51 | def preprocess(self, rgb_image): 52 | rgb_image = self.to_tensor(self.resize(rgb_image)) 53 | return rgb_image[None, ...].cuda() 54 | 55 | def getitem(self, index): 56 | filename = self.filename_list[index] 57 | file_path = os.path.join(self.seq_dir, filename) 58 | rgb_image = self.loader(file_path) 59 | rgb_image = self.preprocess(rgb_image) 60 | 61 | return rgb_image 62 | 63 | 64 | def initialize_models(config_dict): 65 | models = {} 66 | models['encoder'] = ResnetEncoder(num_layers=18, pretrained=False) 67 | models['depth'] = DepthDecoder(num_ch_enc=models['encoder'].num_ch_enc) 68 | 69 | if config_dict.model.type not in posenet_type_dict: 70 | raise ValueError("{} is not a valid type of posenet".format(config_dict.model.type)) 71 | 72 | models['pose'] = posenet_type_dict[config_dict.model.type]() 73 | 74 | state_dict = torch.load(Path(config_dict.model.path) / 'encoder.pth') 75 | models['encoder'].load_state_dict({k : v for k, v in state_dict.items() if k in models['encoder'].state_dict()}) 76 | models['depth'].load_state_dict(torch.load(Path(config_dict.model.path) / 'depth.pth')) 77 | models['pose'].load_state_dict(torch.load(Path(config_dict.model.path) / 'pose.pth')) 78 | 79 | models['encoder'] = models['encoder'].cuda().eval() 80 | models['depth'] = models['depth'].cuda().eval() 81 | models['pose'] = models['pose'].cuda().eval() 82 | 83 | return models 84 | 85 | 86 | def save_traj(file_path: str, pose_list: list) -> None: 87 | pose_list = [p[0:3, 0:4].reshape((12,)) for p in pose_list] 88 | 89 | with open(file_path, 'w') as f: 90 | np.savetxt(f, pose_list, delimiter=' ') 91 | 92 | print('Trajectory saved to {}.'.format(file_path)) 93 | 94 | 95 | def main(): 96 | parser = argparse.ArgumentParser() 97 | parser.add_argument('-c', '--config', type=Path, default='configs/run_odometry.yaml') 98 | args = parser.parse_args() 99 | 100 | with open(args.config, 'r') as f: 101 | config_dict = yaml.load(f, yaml.FullLoader) 102 | config_dict = edict(config_dict) 103 | 104 | odom_dataset = KIITIOdomDataset(config_dict) 105 | models = initialize_models(config_dict) 106 | seq_name = config_dict.data.path.split('/')[-1] 107 | create_dir_if_not_exist(config_dict.output.path) 108 | 109 | traj = [np.eye(4)] 110 | prev_rgb = None 111 | for i in tqdm(range(len(odom_dataset))): 112 | cur_rgb = odom_dataset.getitem(i) 113 | cur_disp = models['depth'](models['encoder'](cur_rgb))['disp', 0] 114 | 115 | if prev_rgb is not None: 116 | rgb_inputs = torch.cat([prev_rgb, cur_rgb], dim=1) 117 | disp_inputs = torch.cat([prev_disp.repeat(1, 3, 1, 1), cur_disp.repeat(1, 3, 1, 1)], dim=1) 118 | 119 | with torch.no_grad(): 120 | euler_angle, translation = models['pose']([rgb_inputs, disp_inputs]) 121 | T = from_euler_t(euler_angle, translation, invert=True) 122 | traj.append(np.dot(traj[-1], T.cpu().numpy()[0])) 123 | 124 | prev_rgb = cur_rgb 125 | prev_disp = cur_disp 126 | 127 | save_traj(os.path.join(config_dict.output.path, '{}.txt'.format(seq_name)), traj) 128 | 129 | if __name__ == '__main__': 130 | main() 131 | --------------------------------------------------------------------------------