├── README.md
└── Resource
    ├── FIGS
        ├── 1111.md
        └── VOT18-22.png
    ├── Light-weight-Tracking.md
    ├── Online-Report.md
    ├── SOT-Paper-Short-Summary.md
    ├── SOT-Papers
        ├── SOT-Papers-1981-2016.md
        ├── SOT-Papers-2016-2020.md
        └── SOT-Papers-2021-2025.md
    ├── SOTBenchmark.md
    ├── Tracking-Framework.md
    ├── Tracking-Survey.md
    └── VOT-Winner.md


/README.md:
--------------------------------------------------------------------------------
 1 | # SOT-Learning
 2 | 
 3 | ## 内容
 4 | 
 5 | 
 6 | 
 7 | ## 资源
 8 | 
 9 |   * **[SOT代码框架](https://github.com/wangdongdut/SOT-Learning/blob/main/Resource/Tracking-Framework.md)**
10 |   * **[SOT综述文章](https://github.com/wangdongdut/SOT-Learning/blob/main/Resource/Tracking-Survey.md)**
11 |   * **[历届VOT冠军算法](https://github.com/wangdongdut/SOT-Learning/blob/main/Resource/VOT-Winner.md)**
12 |   * **[顶会/刊SOT算法速览](https://github.com/wangdongdut/SOT-Learning/blob/main/Resource/SOT-Paper-Short-Summary.md)**
13 |   * **[在线学习/报告资源](https://github.com/wangdongdut/SOT-Learning/blob/main/Resource/Online-Report.md)**
14 | 


--------------------------------------------------------------------------------
/Resource/FIGS/1111.md:
--------------------------------------------------------------------------------
1 | 
2 | 


--------------------------------------------------------------------------------
/Resource/FIGS/VOT18-22.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/wangdongdut/SOT-Learning/93a04ae636219e3bc1ba722fe9e207a78d0604dd/Resource/FIGS/VOT18-22.png


--------------------------------------------------------------------------------
/Resource/Light-weight-Tracking.md:
--------------------------------------------------------------------------------
 1 | # Light-weight Tracking
 2 | 
 3 | * **SiamFC:** Luca Bertinetto, Jack Valmadre, João F. Henriques, Andrea Vedaldi, Philip H. S. Torr. Fully-Convolutional Siamese Networks for Object Tracking. ECCVW, 2016. [[Paper](https://arxiv.org/pdf/1606.09549.pdf)][[Project](https://www.robots.ox.ac.uk/~luca/siamese-fc.html)][[PySOTCode](https://github.com/STVIR/pysot)] 
 4 | * **ECO-HC:** Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, Michael Felsberg. ECO: Efficient Convolution Operators for Tracking. CVPR, 2017. [[Paper](https://arxiv.org/pdf/2103.17154.pdf)][[Code](https://github.com/visionml/pytracking)]
 5 | * **LightTrack:** Bin Yan, Houwen Peng, Kan Wu, Dong Wang, Jianlong Fu, Huchuan Lu. LightTrack: Finding Lightweight Neural Networks for Object Tracking via One-Shot Architecture Search. CVPR, 2021. [[Paper](https://arxiv.org/pdf/2104.14545.pdf)][[Code](https://github.com/researchmm/LightTrack)] 
 6 | * **STARK-Lightning:** Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, Huchuan Lu. Learning Spatio-Temporal Transformer for Visual Tracking. ICCV, 2021. [[Paper](https://arxiv.org/pdf/2103.17154.pdf)][[Code](https://github.com/researchmm/Stark)]
 7 | * **FEAR:** Vasyl Borsuk, Roman Vei, Orest Kupyn, Tetiana Martyniuk, Igor Krashenyi, Jiři Matas. FEAR: Fast, Efficient, Accurate and Robust Visual Tracker. ECCV, 2022. [[Paper](https://arxiv.org/pdf/2112.07957.pdf)][[Code](https://github.com/PinataFarms/FEARTracker)] 
 8 | * **HCAT:** Xin Chen, Dong Wang, Dongdong Li, Huchuan Lu. Efficient Visual Tracking via Hierarchical Cross-Attention Transformer. ECCVW, 2022. [[Paper](https://arxiv.org/pdf/2203.13537.pdf)][[Code](https://github.com/chenxin-dlut/HCAT)] 
 9 | * **E.T.Track:** Philippe Blatter, Menelaos Kanakis, Martin Danelljan, Luc Van Gool. Efficient Visual Tracking with Exemplar Transformers. WACV, 2023. [[Paper](https://arxiv.org/pdf/2112.09686.pdf)][[Code](https://github.com/pblatter/ettrack)] 
10 | * **HiT:** Ben Kang, Xin Chen, Dong Wang, Houwen Peng, Huchuan Lu. Exploring Lightweight Hierarchical Vision Transformers for Efficient Visual Tracking. ICCV, 2023. [[Paper](https://arxiv.org/abs/2308.06904)][[Code](https://github.com/kangben258/HiT)]
11 |  
12 | * **Others:** 
13 |   * **NanoTrack:** https://github.com/HonglinChu/SiamTrackers/tree/master/NanoTrack
14 |   * **让视觉目标跟踪在移动端起飞[YaqiLYU]：** [[一](https://zhuanlan.zhihu.com/p/416413600)], [[二](https://zhuanlan.zhihu.com/p/416754498)], [[三](https://zhuanlan.zhihu.com/p/419900331)]
15 | 
16 | * **Benchmark Results**
17 | 
18 |      | Tracker              | LaSOT (AUC)   | TrackingNet (AUC) | GOT-10K (AO)  |
19 |      |:-----------          |:----------------:|:----------------:|:----------------:|
20 |      | HCAT                 | 59.1  | 76.6 | 65.3 |        
21 |      | E.T.Track            | 59.1  | 70.6 | - |      
22 |      | FEAR                 | 57.9 | - | 64.5 |  
23 |      | STARK-Lightning      | 58.6| - | - |  
24 |      | LightTrack           | 53.8 | 72.5 | 61.1 |      
25 |      | ECO-HC               | 32.4 | 55.4 | 31.6 |  
26 |      | SiamFC               | - | - | - | 
27 | 
28 | <!-- 
29 | 
30 | # Light-weight Transformer Backbone
31 | 
32 | * **DeLighT:** Sachin Mehta, Marjan Ghazvininejad, Srinivasan Iyer, Luke Zettlemoyer, Hannaneh Hajishirzi. DeLighT: Deep and Light-weight Transformer. ICLR, 2021. 
33 |  [[Paper](https://arxiv.org/pdf/2008.00623.pdf)][[Code](https://github.com/sacmehta/delight)]
34 | * **LeViT:** Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, Matthijs Douze. LeViT: a Vision Transformer in ConvNet's Clothing for Faster Inference. ICCV, 2021. [[Paper](https://arxiv.org/pdf/2104.01136.pdf)][[Code](https://github.com/apple/ml-cvnets)]
35 | *  **LVT:** Chenglin Yang, Yilin Wang, Jianming Zhang, He Zhang, Zijun Wei, Zhe Lin, Alan Yuille. Lite Vision Transformer with Enhanced Self-Attention. CVPR, 2022. [[Paper](https://arxiv.org/pdf/2112.10809.pdf)][[Code](https://github.com/Chenglin-Yang/LVT)] 
36 | * **Mobile-Former:** Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, Zicheng Liu. Mobile-Former: Bridging MobileNet and Transformer. CVPR, 2022. [[Paper](https://arxiv.org/pdf/2108.05895.pdf)][[Code](https://github.com/apple/ml-cvnets)] 
37 | * **MobileViT:** 
38 |   * **MobileViT:** Sachin Mehta, Mohammad Rastegari. MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer. ICLR, 2022. [[Paper](https://arxiv.org/pdf/2110.02178v2.pdf)][[Code](https://github.com/apple/ml-cvnets)]
39 |   * **MobileViTv2:** Sachin Mehta, Mohammad Rastegari. Separable Self-attention for Mobile Vision Transformers. Arxiv, 2022. [[Paper](https://arxiv.org/pdf/2206.02680.pdf)][[Code](https://github.com/apple/ml-cvnets)]
40 | * **EdgeViT:** Junting Pan, Adrian Bulat, Fuwen Tan, Xiatian Zhu, Lukasz Dudziak, Hongsheng Li, Georgios Tzimiropoulos, Brais Martinez. EdgeViTs: Competing Light-weight CNNs on Mobile Devices with Vision Transformers. ECCV, 2022. [[Paper](https://arxiv.org/pdf/2205.03436.pdf)][[Code](https://github.com/saic-fi/edgevit)]
41 | * **TinyViT:** Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, Lu Yuan. TinyViT: Fast Pretraining Distillation for Small Vision Transformers. ECCV, 2022. [[Paper](https://arxiv.org/pdf/2207.10666.pdf)][[Code](https://github.com/microsoft/Cream/tree/main/TinyViT)]
42 | * **ParC-Net:** Haokui Zhang, Wenze Hu, Xiaoyu Wang. ParC-Net: Position Aware Circular Convolution with Merits from ConvNets and Transformer. ECCV, 2022. [[Paper](https://arxiv.org/pdf/2203.03952v5.pdf)][[Code](https://github.com/hkzhang91/ParC-Net)]
43 | * **SReT:** Zhiqiang Shen, Zechun Liu, Eric Xing. Sliced Recursive Transformer. ECCV, 2022. [[Paper](https://arxiv.org/pdf/2111.05297.pdf)][[Code](https://github.com/szq0214/SReT)] 
44 | * **LightViT:** Tao Huang, Lang Huang, Shan You, Fei Wang, Chen Qian, Chang Xu. LightViT: Towards Light-Weight Convolution-Free Vision Transformers. Arxiv, 2022. [[Paper](https://arxiv.org/pdf/2207.05557.pdf)][[Code](https://github.com/hunto/LightViT)] 
45 | * **EdgeNeXt:**  Muhammad Maaz, Abdelrahman Shaker, Hisham Cholakkal, Salman Khan, Syed Waqas Zamir, Rao Muhammad Anwer, Fahad Shahbaz Khan. EdgeNeXt: Efficiently Amalgamated CNN-Transformer Architecture for Mobile Vision Applications. Arxiv, 2022. [[Paper](https://arxiv.org/pdf/2206.10589.pdf)][[Code](https://github.com/mmaaz60/EdgeNeXt)]
46 | * **EfficientFormer:** Yanyu Li, Geng Yuan, Yang Wen, Eric Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, Jian Ren. EfficientFormer: Vision Transformers at MobileNet Speed. Arxiv, 2022. [[Paper](https://arxiv.org/pdf/2206.01191.pdf)][[Code](https://github.com/snap-research/EfficientFormer)] 
47 | 
48 | # Light-weight CNN Backbone
49 | 
50 | * **MobileNet:**
51 |   * **MobileNetV1:** Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, Hartwig Adam. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications. Arxiv, 2017. [[Paper](https://arxiv.org/pdf/1704.04861.pdf)]
52 |   * **MobileNetV2:** Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, Liang-Chieh Chen. MobileNetV2: Inverted Residuals and Linear Bottlenecks. CVPR, 2018. [[Paper](https://arxiv.org/pdf/1801.04381.pdf)]
53 |   * **MobileNetV3:** Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, Quoc V. Le, Hartwig Adam. Searching for MobileNetV3. ICCV, 2019. [[Paper](https://arxiv.org/pdf/1905.02244.pdf)]
54 | * **ShuffleNet:**
55 |   * Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, Jian Sun. ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices. CVPR, 2018. [[Paper](https://arxiv.org/pdf/1707.01083.pdf)] 
56 |   * Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, Jian Sun. ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design. ECCV, 2018. [[Paper](https://arxiv.org/pdf/1807.11164.pdf)] 
57 | 
58 | # Reference
59 | * **Ultimate-Awesome-Transformer-Attention:** https://github.com/cmhungsteve/Awesome-Transformer-Attention
60 | * **Efficient-AI-Backbones:** https://github.com/huawei-noah/Efficient-AI-Backbones
61 | * **轻量级网络综述—主干网络篇:** https://mp.weixin.qq.com/s/-LAfyLA3U4C27thFlRgDhw
62 | * **Efficient Transformers: A Survey** https://arxiv.org/abs/2009.06732
63 | 
64 | -->
65 | 


--------------------------------------------------------------------------------
/Resource/Online-Report.md:
--------------------------------------------------------------------------------
 1 | * **在线学习/报告资源**
 2 | 
 3 |      | **视频/博客**              | 内容摘要    |
 4 |      |:-----------               |:----------------:|
 5 |      |**[极市平台](https://space.bilibili.com/85300886/)(跟踪问题)：[B站视频](https://space.bilibili.com/85300886/search/video?keyword=%E8%B7%9F%E8%B8%AA)** |部分原作者论文报告；公众号：极市平台|
 6 |      |**[VALSE_Webinar](https://space.bilibili.com/562085182/)(跟踪问题)：[B站视频](https://space.bilibili.com/562085182/search/video?keyword=%E8%B7%9F%E8%B8%AA)** |基础知识，前沿报告，论文速览；公众号：VALSE |
 7 |      |**[叫我林林就行](https://space.bilibili.com/209664735/)：[Github](https://github.com/HonglinChu), [B站视频](https://space.bilibili.com/209664735/)** |CF和Siamese系列工作介绍复现；轻量级跟踪器NanoTrack|
 8 |      |**[东大阿德](https://space.bilibili.com/6794832/)：[单目标跟踪零基础代码入门-B站视频](https://space.bilibili.com/6794832/channel/collectiondetail?sid=400776)** |SiamFC, SiamRPN, pytracking系列代码讲解；RGBT跟踪介绍|
 9 |      |**[深度大脑](https://space.bilibili.com/9456738)：[从零开始的单目标跟踪算法世界-B站视频](https://www.bilibili.com/video/BV1WK4y1C7JG)** |pytracking系列代码讲解|
10 |      |**[AITracker](https://space.bilibili.com/8948069)：[B站视频](https://space.bilibili.com/8948069)** |系列讲座，经验分享；公众号：目标跟踪与检测基础前沿|
11 |      |**[人工智能前沿讲习](https://bbs.sffai.com/)(跟踪问题)：[B站视频](https://space.bilibili.com/388690539/search/video?keyword=%E8%B7%9F%E8%B8%AA)** |系列讲座，经验分享；公众号：人工智能前沿讲习|
12 |      |**[YaqiLYU(知乎)](https://www.zhihu.com/people/YaqiLYU)：[深度学习和目标跟踪(知乎专栏)](https://www.zhihu.com/column/DCF-tracking)** |算法总结感悟，移动端跟踪|
13 |      |**[NTRのSAO年(B站)](https://space.bilibili.com/5567932)：[B站专栏](https://space.bilibili.com/5567932/article)** |部分论文速览|
14 | 


--------------------------------------------------------------------------------
/Resource/SOT-Paper-Short-Summary.md:
--------------------------------------------------------------------------------
 1 | *  **SOT-SOTA：**
 2 |     * LaSOT：https://paperswithcode.com/sota/visual-object-tracking-on-lasot
 3 |     * TrackingNet：
 4 |       * http://eval.tracking-net.org/featured-challenges/39/leaderboard/42
 5 |       * https://paperswithcode.com/sota/visual-object-tracking-on-trackingnet
 6 |     * Got-10K：http://got-10k.aitestunion.com/leaderboard
 7 |     * OTB：https://paperswithcode.com/sota/visual-object-tracking-on-otb-2015
 8 | 
 9 | * **顶会/刊SOT算法速览：历年视觉三大会（CVPR，ICCV，ECCV），其他会议/期刊/Workshop上部分论文**
10 |   
11 |   * **[2021年-2025年](https://github.com/wangdongdut/SOT-Learning/blob/main/Resource/SOT-Papers/SOT-Papers-2021-2025.md)**
12 |   
13 |   * **[2016年-2020年](https://github.com/wangdongdut/SOT-Learning/blob/main/Resource/SOT-Papers/SOT-Papers-2016-2020.md)**
14 |   
15 |   * **[1981年-2016年](https://github.com/wangdongdut/SOT-Learning/blob/main/Resource/SOT-Papers/SOT-Papers-1981-2016.md)**
16 | 


--------------------------------------------------------------------------------
/Resource/SOT-Papers/SOT-Papers-1981-2016.md:
--------------------------------------------------------------------------------
1 |  * **2021年SOT Papers：**
2 | 
3 |     | **来源**              | **官方/知名缩写**     | **题目/链接/代码/Bibtex**    | **简介**    |
4 |     |:-----------           |:----------------|:----------------|:----------------|
5 |     | ECCV2022| Unicorn | Towards Grand Unification of Object Tracking </br> [[Paper]()][[Code]()][[Bibtex]()]|
6 |     | | | |
7 | 


--------------------------------------------------------------------------------
/Resource/SOT-Papers/SOT-Papers-2016-2020.md:
--------------------------------------------------------------------------------
 1 |  * **2022年SOT Papers：**
 2 | 
 3 |     | **来源**              | **官方/知名缩写**     | **题目/链接/代码/Bibtex**    | **简介**    |
 4 |     |:-----------           |:----------------|:----------------|:----------------|
 5 |     | ECCV2022| Unicorn | Towards Grand Unification of Object Tracking </br> [[Paper]()][[Code]()]|
 6 |     | | | |
 7 |     | | | |
 8 |     | | | |
 9 |     | CVPR2022 |ToMP  | Transforming Model Prediction for Tracking  </br> [[Paper](https://openaccess.thecvf.com/content/CVPR2022/html/Mayer_Transforming_Model_Prediction_for_Tracking_CVPR_2022_paper.html)][[Code](https://github.com/visionml/pytracking)] |
10 | 


--------------------------------------------------------------------------------
/Resource/SOT-Papers/SOT-Papers-2021-2025.md:
--------------------------------------------------------------------------------
 1 |  * **2022年SOT Papers：**
 2 | 
 3 |     | **来源**              | **官方/知名缩写**     | **题目/链接/代码**    | **简介**    |
 4 |     |:-----------           |:----------------|:----------------|:----------------|
 5 |     | ECCV2022| Unicorn | Towards Grand Unification of Object Tracking </br> [[Paper]()][[Code]()]|
 6 |     | | | |
 7 |     | | | |
 8 |     | | | |
 9 |     | CVPR2022 | ToMP | Transforming Model Prediction for Tracking  </br> [[Paper](https://openaccess.thecvf.com/content/CVPR2022/html/Mayer_Transforming_Model_Prediction_for_Tracking_CVPR_2022_paper.html)][[Code](https://github.com/visionml/pytracking)] | |
10 |     | CVPR2022 | ULAST | Unsupervised Learning of Accurate Siamese Tracking </br> [[Paper](https://openaccess.thecvf.com/content/CVPR2022/html/Shen_Unsupervised_Learning_of_Accurate_Siamese_Tracking_CVPR_2022_paper.html)][[Code](https://github.com/FlorinShum/ULAST)] |  |
11 |     | CVPR2022 | UTT | Unsupervised Learning of Accurate Siamese Tracking </br> [[Paper](https://openaccess.thecvf.com/content/CVPR2022/html/Ma_Unified_Transformer_Tracker_for_Object_Tracking_CVPR_2022_paper.html)][[Code](https://github.com/Flowerfan/Trackron)] |  |
12 |     | CVPR2022 | STNet | Spiking Transformers for Event-Based Single Object Tracking </br> [[Paper](https://openaccess.thecvf.com/content/CVPR2022/html/Shen_Unsupervised_Learning_of_Accurate_Siamese_Tracking_CVPR_2022_paper.html)][[Code](https://github.com/Jee-King/CVPR2022_STNet)] |  |
13 |     | **CVPR2022** | **UDAT** | **Unsupervised Domain Adaptation for Nighttime Aerial Tracking** </br> Fan Ma, Mike Zheng Shou, Linchao Zhu, Haoqi Fan, Yilei Xu, Yi Yang, Zhicheng Yan </br> [[Paper](https://openaccess.thecvf.com/content/CVPR2022/html/Ye_Unsupervised_Domain_Adaptation_for_Nighttime_Aerial_Tracking_CVPR_2022_paper.html)][[Code](https://github.com/vision4robotics/UDAT)] |  |
14 |     | **CVPR2022** | **TCTrack** | **TCTrack: Temporal Contexts for Aerial Tracking** </br> Ziang Cao, Ziyuan Huang, Liang Pan, Shiwei Zhang, Ziwei Liu, Changhong Fu </br> [[Paper](https://openaccess.thecvf.com/content/CVPR2022/html/Cao_TCTrack_Temporal_Contexts_for_Aerial_Tracking_CVPR_2022_paper.html)][[Code](https://github.com/vision4robotics/TCTrack)] |  |
15 |     | **CVPR2022** | **SBT** | **Correlation-Aware Deep Tracking** </br> Fei Xie, Chunyu Wang, Guangting Wang, Yue Cao, Wankou Yang, Wenjun Zeng </br> [[Paper](https://openaccess.thecvf.com/content/CVPR2022/html/Xie_Correlation-Aware_Deep_Tracking_CVPR_2022_paper.html)][[Code](https://github.com/phiphiphi31/SuperSBT)] |  |
16 |     | **CVPR2022** | **CSWinTT** | **Transformer Tracking With Cyclic Shifting Window Attention** </br> Zikai Song, Junqing Yu, Yi-Ping Phoebe Chen, Wei Yang </br> [[Paper](https://openaccess.thecvf.com/content/CVPR2022/html/Song_Transformer_Tracking_With_Cyclic_Shifting_Window_Attention_CVPR_2022_paper.html)][[Code](https://github.com/SkyeSong38/CSWinTT)] |  |
17 |     | **CVPR2022** | **GTELT** | **Global Tracking via Ensemble of Local Trackers** </br> Zikun Zhou, Jianqiu Chen, Wenjie Pei, Kaige Mao, Hongpeng Wang, Zhenyu He </br> [[Paper](https://openaccess.thecvf.com/content/CVPR2022/html/Zhou_Global_Tracking_via_Ensemble_of_Local_Trackers_CVPR_2022_paper.html)][[Code](https://github.com/ZikunZhou/GTELT)] |  |
18 |     | **CVPR2022** | **MixFormer** | **MixFormer: End-to-End Tracking With Iterative Mixed Attention** </br> Yutao Cui, Cheng Jiang, Limin Wang, Gangshan Wu </br> [[Paper](https://openaccess.thecvf.com/content/CVPR2022/html/Cui_MixFormer_End-to-End_Tracking_With_Iterative_Mixed_Attention_CVPR_2022_paper.html)][[Code](https://github.com/MCG-NJU/MixFormer)] |  |
19 |      | **CVPR2022** | **RBO** | **Ranking-Based Siamese Visual Tracking** </br> Feng Tang, Qiang Ling </br> [[Paper](https://openaccess.thecvf.com/content/CVPR2022/html/Cui_MixFormer_End-to-End_Tracking_With_Iterative_Mixed_Attention_CVPR_2022_paper.html)][[Code](https://github.com/sansanfree/RBO)] |  |   
20 |      | **CVPR2022** | **VTUAV** | **Visible-Thermal UAV Tracking: A Large-Scale Benchmark and New Baseline** </br> Pengyu Zhang, Jie Zhao, Dong Wang, Huchuan Lu, Xiang Ruan </br> [[Paper](https://openaccess.thecvf.com/content/CVPR2022/html/Zhang_Visible-Thermal_UAV_Tracking_A_Large-Scale_Benchmark_and_New_Baseline_CVPR_2022_paper.html)][[Code](https://zhang-pengyu.github.io/DUT-VTUAV/)] |  | 
21 | 


--------------------------------------------------------------------------------
/Resource/SOTBenchmark.md:
--------------------------------------------------------------------------------
1 | 
2 | 
3 | https://machinelearning.uniud.it/datasets/trek150/
4 | 


--------------------------------------------------------------------------------
/Resource/Tracking-Framework.md:
--------------------------------------------------------------------------------
 1 | * **SOT代码框架**
 2 | 
 3 |      | **框架**              | **链接**    | **简介**    |
 4 |      |:-----------           |:----------------|:----------------|
 5 |      | **pysot**             | https://github.com/STVIR/pysot             | STVIR组提供的跟踪框架，主要包括SiamFC，SiamRPN，DaSiamRPN，SiamRPN++，SiamMask等算法。|     
 6 |      | **pytracking**        | https://github.com/visionml/pytracking     | visionml组提供的跟踪框架，主要包括pyECO, ATOM, Dimp, PrDimp, KeepTrack等算法。|     
 7 |      | **TracKit**           | https://github.com/researchmm/TracKit      | MSRA-MM组提供的跟踪框架，主要包括SiamDW，Ocean跟踪算法。|
 8 |      | **mmtracking**        | https://github.com/open-mmlab/mmtracking   | Open-mmlab Tracking相关库，包括视频目标检测，单目标跟踪，多目标跟踪，视频实例分割。其中单目标跟踪算法支持SiameseRPN++(CVPR2019)，STARK(ICCV2021)，PrDiMP(CVPR2020)等算法。和Open-mmlab一脉相承，未来有很大潜力。 |
 9 |      | **got-10k-toolkit**   | https://github.com/got-10k/toolkit   | Got-10k数据集网站提供的评估工具，方便大量跟踪数据库上的实验测评，如OTB(2013/2015), VOT(2013~2018), DTB70, TColor128, NfS(30/240 fps), UAV(123/20L), LaSOT, TrackingNet等。 |
10 |      | **opencv**            | https://opencv.org/   | 包括CamShift，MIL，GOTURN，DaSiamRPN，KCF, TLD等代码或API。 |
11 |      | **PaddlePaddle**<br /> PaddleCV/tracking | https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/tracking   | 包括SiamFC, SiamRPN, SiamMask, ATOM等算法。 |
12 | 


--------------------------------------------------------------------------------
/Resource/Tracking-Survey.md:
--------------------------------------------------------------------------------
 1 | * **SOT算法综述**     
 2 |    
 3 |      | **论文**              | **链接**    | **内容摘要**    |
 4 |      |:-----------           |:----------------:|:----------------|
 5 |      | ***Visual Object Tracking with Discriminative Filters and Siamese Networks: A Survey and Outlook.*** <br /> **IEEE TPAMI, 2022.** Sajid Javed, Martin Danelljan, Fahad Shahbaz Khan, et al. | [[Paper](https://arxiv.org/abs/2112.02838)]   |  **:star2:非常详尽的判决性滤波和深度孪生网络方面的综述**，大量的实验评估。SiamR-CNN基本上算是深度孪生卷积网络类跟踪算法的精度巅峰，但于2021年被简单的Transformer跟踪算法超越。*2021年开始，SOT全面进入Transformer时代。* | 
 6 |      | ***Deep Learning for Visual Tracking: A Comprehensive Survey.*** <br /> **IEEE TITS, 2022.** Seyed Mojtaba Marvasti-Zadeh, Li Cheng, Hossein Ghanei-Yakhdan, Shohreh Kasae.     |  [[Paper](https://ieeexplore.ieee.org/document/9339950)]   |  **:star2:非常详尽的深度学习类跟踪算法综述**， 从网络结构，网络利用，网络训练，网络目标，网络输出，深度学习与相关滤波，长时跟踪，在线更新跟踪等多个维度详尽地综述了2013-2020年的深度学习类跟踪算法。<br /> *2019-2020时的深度学习类跟踪相关已显著超越相关滤波结合深度特征类算法，SOT已彻底全面步入深度学习时代。* |   
 7 |      | ***Handcrafted and Deep Trackers: Recent Visual Object Tracking Approaches and Trends.*** <br />  **ACM CS, 2019.** Mustansar Fiaz, Arif Mahmood, Sajid Javed, Soon Ki Jung.     |  [[Paper](https://dl.acm.org/doi/abs/10.1145/3309665)]   |  主要综述了2018年之前的相关滤波类算法和非相关滤波类算法，对2015-2018年提出的相关滤波类算法综述较为详尽。 *深度学习类跟踪算法更推荐看《Deep Learning for Visual Tracking: A Comprehensive Survey》*  |  
 8 |      | ***Deep Visual Tracking: Review and Experimental Comparison.*** <br />  **PR, 2018.** Peixia Li, Dong Wang, Lijun Wang, Huchuan Lu.    |  [[Paper](https://www.sciencedirect.com/science/article/abs/pii/S0031320317304612)]   |   主要综述了早期深度学习跟踪算法，从网络结构，网络形式，网络训练三个角度进行了细致的分析。当时，在端到端深度跟踪算法和传统跟踪算法结合深度特征还平分秋色；2019年后，端到端深度跟踪算法占据绝对统治地位。*深度学习类跟踪算法更推荐看《Deep Learning for Visual Tracking: A Comprehensive Survey》*  |  
 9 |      | ***Handcrafted and Deep Trackers: Recent Visual Object Tracking Approaches and Trends.*** <br />  **ACM CS, 2019.** Mustansar Fiaz, Arif Mahmood, Sajid Javed, Soon Ki Jung.     |  [[Paper](https://dl.acm.org/doi/abs/10.1145/3309665)]   |  主要综述了2018年之前的相关滤波类算法和非相关滤波类算法，对2015-2018年提出的相关滤波类算法综述较为详尽。<br />  *深度学习类跟踪算法更推荐看《Deep Learning for Visual Tracking: A Comprehensive Survey》*  |  
10 |      | ***Deep Visual Tracking: Review and Experimental Comparison.*** <br />  **PR, 2018.** Peixia Li, Dong Wang, Lijun Wang, Huchuan Lu.    |  [[Paper](https://www.sciencedirect.com/science/article/abs/pii/S0031320317304612)]   |   主要综述了早期深度学习跟踪算法，从网络结构，网络形式，网络训练三个角度进行了细致的分析。当时，在端到端深度跟踪算法和传统跟踪算法结合深度特征还平分秋色；2019年后，端到端深度跟踪算法占据绝对统治地位。<br /> *深度学习类跟踪算法更推荐看《Deep Learning for Visual Tracking: A Comprehensive Survey》*  |  
11 |      | ***Object Tracking Benchmark.*** <br />  **IEEE TPAMI, 2015.** Yi Wu, Jongwoo Lim, Ming-Hsuan Yang. |  [[Paper](https://ieeexplore.ieee.org/document/7001050)] <br /> [[Website](http://cvlab.hanyang.ac.kr/tracker_benchmark/)]   | **跟踪领域最经典Benchmark：OTB-100**，同样也是2013年及之前跟踪算法很好的总结，推动了跟踪算法评测的标准化，提出的**Precision和Success**指标现在仍是主流评测指标。评测了29个代表性的Trackers，但排名未进前5的CSK算法在结合了HOG特征之后，成为了2015-2019年最主要的主流算法，即相关滤波算法。| 
12 |      | ***Understanding and Diagnosing Visual Tracking Systems.*** <br /> **ICCV, 2015.**  Naiyan Wang, Jianping Shi, Dit-Yan Yeung, Jiaya Jia.   | [[Paper](https://ieeexplore.ieee.org/document/7410712)] <br />  [[Website](http://www.winsty.net/tracker_diagnose.html)]  | 一份非常值得读的早期跟踪系统解析报告，重要结论：跟踪器性能**特征为王**，运动模型不那么重要（这个其实有数据集偏见在里面）。对现在SOT等相关领域研究也有启发意义。如果发现**所选用的特征**落后时代了，那么把外观模型（公式或结构）搞的再花哨也是意义不大的。|  
13 |      | ***Visual Tracking: An Experimental Survey.*** <br /> **IEEE TPAMI, 2014.**  Arnold W. M. Smeulders, Dung Manh Chu, Rita Cucchiara, et al.   | [[Paper](https://ieeexplore.ieee.org/document/6671560)]    | 早期的跟踪评估综述，主要总结和评估了2012年前的知名算法。提出了有315个视频的评测数据集ALOV++，但由于在CVPR2013和TPAMI2015上发表和扩充的OTB数据库和评测标准更受研究者欢迎，ALOV++现在鲜有人用，而OTB已成为业内标准之一。   |   
14 | 


--------------------------------------------------------------------------------
/Resource/VOT-Winner.md:
--------------------------------------------------------------------------------
 1 | * **历年[VOT](https://www.votchallenge.net/)冠军算法（VOT2013-VOT2022）**
 2 |     
 3 |      * ST：短时跟踪，也被称为main challenge；RT：实时跟踪，在ST设置基础上要求算法要能跟上视频流；TIR：热红外；RGB-T：RGB彩色+热红外；LT：强调目标消失再出现，需要给出跟踪结果的置信值；RGBD：RGB彩色+深度
 4 |      * VOT竞赛冠军方案中主要涉及的跟踪模型：STARK(ICCV2021)，TransT(CVPR2021)，AlphaRefine(CVPR2021)，RPT(ICCVW2021)，Meta-Updater(CVPR2020), Dimp(ICCV2019), ATOM(CVPR2021), SiamMask(CVPR2019), SiamRPN++(CVPR2019), SiamDW(CVPR2019)，SPLT(ICCV2019)，SiamRPN(CVPR2018)，RT-MDNet(ECCV2018)，DSLT(PRL2018)，MDNet(CVPR2016)，SiamFC(ECCVW2016)，EBT(CVPR2016)，KCF(TPAMI2015)，PST(ICCV2015)，DSST(BMVC2014)，Struck(CVPR2011)。
 5 |      
 6 |      | **年份**              | **赛道**    | **冠军简介**    |
 7 |      |:-----------           |:----------------|:----------------|    
 8 |      | [**VOT2022**](https://www.votchallenge.net/vot2021/)  | RGBD | **STARK_RGBD**(STARK for the RGBD challenge)：This is a method combining STARK [[66](https://openaccess.thecvf.com/content/ICCV2021/html/Yan_Learning_Spatio-Temporal_Transformer_for_Visual_Tracking_ICCV_2021_paper.html)] and DiMPsuper [[2](https://ieeexplore.ieee.org/document/9010649)]. STARK is a powerful transformer-based tracker with a siamese structure. We first change the backbone of STARK into DeiT [[57](https://proceedings.mlr.press/v139/touvron21a.html)]. The transformer-based backbone DeiT strengthens the feature of STARK. We notice the STARK method is not good at handling the appearance change of the target. To this end, we combine the DeiT-strengthened STARK with the DiMPsuper tracker, a powerful online updating tracker. Specifically, when the STARK tracker’s confidence is low or the prediction of STARK suddenly strays away, the DiMPsuper takes over the tracking process, providing a steady, appearance adaptive result. When the STARK’s confidence resumes, we switch back to STARK. We also design a refinement module similar to AlphaRefine [[67](https://ieeexplore.ieee.org/document/9577854)] by modifying the search region of STARK. The refinement module is applied to the final output of the whole tracking system for further boosting the quality of box estimation. --From [《VOT2021 Report》](https://prints.vicos.si/publications/400)<br />**核心：[STARK算法](https://openaccess.thecvf.com/content/ICCV2021/html/Yan_Learning_Spatio-Temporal_Transformer_for_Visual_Tracking_ICCV_2021_paper.html)(ICCV2021)，结合[SuperDimp](https://ieeexplore.ieee.org/document/9010649)算法，再利用[AlphaRefine](https://ieeexplore.ieee.org/document/9577854)修正结果。**|    
 9 |      | [**VOT2022**](https://www.votchallenge.net/vot2021/)  | RGBD | **STARK_RGBD**(STARK for the RGBD challenge)：This is a method combining STARK [[66](https://openaccess.thecvf.com/content/ICCV2021/html/Yan_Learning_Spatio-Temporal_Transformer_for_Visual_Tracking_ICCV_2021_paper.html)] and DiMPsuper [[2](https://ieeexplore.ieee.org/document/9010649)]. STARK is a powerful transformer-based tracker with a siamese structure. We first change the backbone of STARK into DeiT [[57](https://proceedings.mlr.press/v139/touvron21a.html)]. The transformer-based backbone DeiT strengthens the feature of STARK. We notice the STARK method is not good at handling the appearance change of the target. To this end, we combine the DeiT-strengthened STARK with the DiMPsuper tracker, a powerful online updating tracker. Specifically, when the STARK tracker’s confidence is low or the prediction of STARK suddenly strays away, the DiMPsuper takes over the tracking process, providing a steady, appearance adaptive result. When the STARK’s confidence resumes, we switch back to STARK. We also design a refinement module similar to AlphaRefine [[67](https://ieeexplore.ieee.org/document/9577854)] by modifying the search region of STARK. The refinement module is applied to the final output of the whole tracking system for further boosting the quality of box estimation. --From [《VOT2021 Report》](https://prints.vicos.si/publications/400)<br />**核心：[STARK算法](https://openaccess.thecvf.com/content/ICCV2021/html/Yan_Learning_Spatio-Temporal_Transformer_for_Visual_Tracking_ICCV_2021_paper.html)(ICCV2021)，结合[SuperDimp](https://ieeexplore.ieee.org/document/9010649)算法，再利用[AlphaRefine](https://ieeexplore.ieee.org/document/9577854)修正结果。**|  
10 |      | [**VOT2022**](https://www.votchallenge.net/vot2021/)  | RGBD | **STARK_RGBD**(STARK for the RGBD challenge)：This is a method combining STARK [[66](https://openaccess.thecvf.com/content/ICCV2021/html/Yan_Learning_Spatio-Temporal_Transformer_for_Visual_Tracking_ICCV_2021_paper.html)] and DiMPsuper [[2](https://ieeexplore.ieee.org/document/9010649)]. STARK is a powerful transformer-based tracker with a siamese structure. We first change the backbone of STARK into DeiT [[57](https://proceedings.mlr.press/v139/touvron21a.html)]. The transformer-based backbone DeiT strengthens the feature of STARK. We notice the STARK method is not good at handling the appearance change of the target. To this end, we combine the DeiT-strengthened STARK with the DiMPsuper tracker, a powerful online updating tracker. Specifically, when the STARK tracker’s confidence is low or the prediction of STARK suddenly strays away, the DiMPsuper takes over the tracking process, providing a steady, appearance adaptive result. When the STARK’s confidence resumes, we switch back to STARK. We also design a refinement module similar to AlphaRefine [[67](https://ieeexplore.ieee.org/document/9577854)] by modifying the search region of STARK. The refinement module is applied to the final output of the whole tracking system for further boosting the quality of box estimation. --From [《VOT2021 Report》](https://prints.vicos.si/publications/400)<br />**核心：[STARK算法](https://openaccess.thecvf.com/content/ICCV2021/html/Yan_Learning_Spatio-Temporal_Transformer_for_Visual_Tracking_ICCV_2021_paper.html)(ICCV2021)，结合[SuperDimp](https://ieeexplore.ieee.org/document/9010649)算法，再利用[AlphaRefine](https://ieeexplore.ieee.org/document/9577854)修正结果。**|   
11 |      | [**VOT2022**](https://www.votchallenge.net/vot2021/)  | RGBD | **STARK_RGBD**(STARK for the RGBD challenge)：This is a method combining STARK [[66](https://openaccess.thecvf.com/content/ICCV2021/html/Yan_Learning_Spatio-Temporal_Transformer_for_Visual_Tracking_ICCV_2021_paper.html)] and DiMPsuper [[2](https://ieeexplore.ieee.org/document/9010649)]. STARK is a powerful transformer-based tracker with a siamese structure. We first change the backbone of STARK into DeiT [[57](https://proceedings.mlr.press/v139/touvron21a.html)]. The transformer-based backbone DeiT strengthens the feature of STARK. We notice the STARK method is not good at handling the appearance change of the target. To this end, we combine the DeiT-strengthened STARK with the DiMPsuper tracker, a powerful online updating tracker. Specifically, when the STARK tracker’s confidence is low or the prediction of STARK suddenly strays away, the DiMPsuper takes over the tracking process, providing a steady, appearance adaptive result. When the STARK’s confidence resumes, we switch back to STARK. We also design a refinement module similar to AlphaRefine [[67](https://ieeexplore.ieee.org/document/9577854)] by modifying the search region of STARK. The refinement module is applied to the final output of the whole tracking system for further boosting the quality of box estimation. --From [《VOT2021 Report》](https://prints.vicos.si/publications/400)<br />**核心：[STARK算法](https://openaccess.thecvf.com/content/ICCV2021/html/Yan_Learning_Spatio-Temporal_Transformer_for_Visual_Tracking_ICCV_2021_paper.html)(ICCV2021)，结合[SuperDimp](https://ieeexplore.ieee.org/document/9010649)算法，再利用[AlphaRefine](https://ieeexplore.ieee.org/document/9577854)修正结果。**|  
12 |      | [**VOT2022**](https://www.votchallenge.net/vot2021/)  | RGBD | **STARK_RGBD**(STARK for the RGBD challenge)：This is a method combining STARK [[66](https://openaccess.thecvf.com/content/ICCV2021/html/Yan_Learning_Spatio-Temporal_Transformer_for_Visual_Tracking_ICCV_2021_paper.html)] and DiMPsuper [[2](https://ieeexplore.ieee.org/document/9010649)]. STARK is a powerful transformer-based tracker with a siamese structure. We first change the backbone of STARK into DeiT [[57](https://proceedings.mlr.press/v139/touvron21a.html)]. The transformer-based backbone DeiT strengthens the feature of STARK. We notice the STARK method is not good at handling the appearance change of the target. To this end, we combine the DeiT-strengthened STARK with the DiMPsuper tracker, a powerful online updating tracker. Specifically, when the STARK tracker’s confidence is low or the prediction of STARK suddenly strays away, the DiMPsuper takes over the tracking process, providing a steady, appearance adaptive result. When the STARK’s confidence resumes, we switch back to STARK. We also design a refinement module similar to AlphaRefine [[67](https://ieeexplore.ieee.org/document/9577854)] by modifying the search region of STARK. The refinement module is applied to the final output of the whole tracking system for further boosting the quality of box estimation. --From [《VOT2021 Report》](https://prints.vicos.si/publications/400)<br />**核心：[STARK算法](https://openaccess.thecvf.com/content/ICCV2021/html/Yan_Learning_Spatio-Temporal_Transformer_for_Visual_Tracking_ICCV_2021_paper.html)(ICCV2021)，结合[SuperDimp](https://ieeexplore.ieee.org/document/9010649)算法，再利用[AlphaRefine](https://ieeexplore.ieee.org/document/9577854)修正结果。**|  
13 |     | [**VOT2022**](https://www.votchallenge.net/vot2021/)  | RGBD | **STARK_RGBD**(STARK for the RGBD challenge)：This is a method combining STARK [[66](https://openaccess.thecvf.com/content/ICCV2021/html/Yan_Learning_Spatio-Temporal_Transformer_for_Visual_Tracking_ICCV_2021_paper.html)] and DiMPsuper [[2](https://ieeexplore.ieee.org/document/9010649)]. STARK is a powerful transformer-based tracker with a siamese structure. We first change the backbone of STARK into DeiT [[57](https://proceedings.mlr.press/v139/touvron21a.html)]. The transformer-based backbone DeiT strengthens the feature of STARK. We notice the STARK method is not good at handling the appearance change of the target. To this end, we combine the DeiT-strengthened STARK with the DiMPsuper tracker, a powerful online updating tracker. Specifically, when the STARK tracker’s confidence is low or the prediction of STARK suddenly strays away, the DiMPsuper takes over the tracking process, providing a steady, appearance adaptive result. When the STARK’s confidence resumes, we switch back to STARK. We also design a refinement module similar to AlphaRefine [[67](https://ieeexplore.ieee.org/document/9577854)] by modifying the search region of STARK. The refinement module is applied to the final output of the whole tracking system for further boosting the quality of box estimation. --From [《VOT2021 Report》](https://prints.vicos.si/publications/400)<br />**核心：[STARK算法](https://openaccess.thecvf.com/content/ICCV2021/html/Yan_Learning_Spatio-Temporal_Transformer_for_Visual_Tracking_ICCV_2021_paper.html)(ICCV2021)，结合[SuperDimp](https://ieeexplore.ieee.org/document/9010649)算法，再利用[AlphaRefine](https://ieeexplore.ieee.org/document/9577854)修正结果。**|  
14 |      | [**VOT2022**](https://www.votchallenge.net/vot2021/)  | RGBD | **STARK_RGBD**(STARK for the RGBD challenge)：This is a method combining STARK [[66](https://openaccess.thecvf.com/content/ICCV2021/html/Yan_Learning_Spatio-Temporal_Transformer_for_Visual_Tracking_ICCV_2021_paper.html)] and DiMPsuper [[2](https://ieeexplore.ieee.org/document/9010649)]. STARK is a powerful transformer-based tracker with a siamese structure. We first change the backbone of STARK into DeiT [[57](https://proceedings.mlr.press/v139/touvron21a.html)]. The transformer-based backbone DeiT strengthens the feature of STARK. We notice the STARK method is not good at handling the appearance change of the target. To this end, we combine the DeiT-strengthened STARK with the DiMPsuper tracker, a powerful online updating tracker. Specifically, when the STARK tracker’s confidence is low or the prediction of STARK suddenly strays away, the DiMPsuper takes over the tracking process, providing a steady, appearance adaptive result. When the STARK’s confidence resumes, we switch back to STARK. We also design a refinement module similar to AlphaRefine [[67](https://ieeexplore.ieee.org/document/9577854)] by modifying the search region of STARK. The refinement module is applied to the final output of the whole tracking system for further boosting the quality of box estimation. --From [《VOT2021 Report》](https://prints.vicos.si/publications/400)<br />**核心：[STARK算法](https://openaccess.thecvf.com/content/ICCV2021/html/Yan_Learning_Spatio-Temporal_Transformer_for_Visual_Tracking_ICCV_2021_paper.html)(ICCV2021)，结合[SuperDimp](https://ieeexplore.ieee.org/document/9010649)算法，再利用[AlphaRefine](https://ieeexplore.ieee.org/document/9577854)修正结果。**|  
15 |      | [**VOT2021**](https://www.votchallenge.net/vot2021/)  | RGBD | **STARK_RGBD**(STARK for the RGBD challenge)：This is a method combining STARK [[66](https://openaccess.thecvf.com/content/ICCV2021/html/Yan_Learning_Spatio-Temporal_Transformer_for_Visual_Tracking_ICCV_2021_paper.html)] and DiMPsuper [[2](https://ieeexplore.ieee.org/document/9010649)]. STARK is a powerful transformer-based tracker with a siamese structure. We first change the backbone of STARK into DeiT [[57](https://proceedings.mlr.press/v139/touvron21a.html)]. The transformer-based backbone DeiT strengthens the feature of STARK. We notice the STARK method is not good at handling the appearance change of the target. To this end, we combine the DeiT-strengthened STARK with the DiMPsuper tracker, a powerful online updating tracker. Specifically, when the STARK tracker’s confidence is low or the prediction of STARK suddenly strays away, the DiMPsuper takes over the tracking process, providing a steady, appearance adaptive result. When the STARK’s confidence resumes, we switch back to STARK. We also design a refinement module similar to AlphaRefine [[67](https://ieeexplore.ieee.org/document/9577854)] by modifying the search region of STARK. The refinement module is applied to the final output of the whole tracking system for further boosting the quality of box estimation. --From [《VOT2021 Report》](https://prints.vicos.si/publications/400)<br />**核心：[STARK算法](https://openaccess.thecvf.com/content/ICCV2021/html/Yan_Learning_Spatio-Temporal_Transformer_for_Visual_Tracking_ICCV_2021_paper.html)(ICCV2021)，结合[SuperDimp](https://ieeexplore.ieee.org/document/9010649)算法，再利用[AlphaRefine](https://ieeexplore.ieee.org/document/9577854)修正结果。**|  
16 |      | [**VOT2021**](https://www.votchallenge.net/vot2021/)  | LT | **mlpLT**(Fusing Complementary Trackers for Long-term Visual Tracking)： The idea behind the mlpLT tracker is to fuse the capabilities of different trackers. In particular, mlpLT implements a strategy that fuses the Stark [[66](https://openaccess.thecvf.com/content/ICCV2021/html/Yan_Learning_Spatio-Temporal_Transformer_for_Visual_Tracking_ICCV_2021_paper.html)] and SuperDiMP [[2](https://ieeexplore.ieee.org/document/9010649)] trackers, which have been selected due to their complementary features. Indeed, the tracker Stark has been selected because of its ability to provide spatially accurate bounding-boxes, and to re-detect the target after disappearances. The meta-updater [[14](https://ieeexplore.ieee.org/document/9156764)] controlled SuperDiMP was chosen due its robustness. The combination of such trackers is managed by a decision strategy based on an online learned target verifier [[50](https://link.springer.com/chapter/10.1007/978-3-030-01225-0_6)]. At every frame, the trackers are run in parallel to predict their target localizations. Such outputs are checked by the verifier which quantifies how good the trackers are following the target. Based on such evaluations, the decision strategy selects which localization to give as output for the current frame. Such an outcome is also employed to correct the tracker that achieved the lowest performance according to the verifier. Additionally strategies such as the computation of adaptive search areas, and the avoidance of wrong target size estimations, have been implemented to the baseline trackers in order to make their localizations more consistent. --From [《VOT2021 Report》](https://prints.vicos.si/publications/400)<br />**核心：[STARK算法](https://openaccess.thecvf.com/content/ICCV2021/html/Yan_Learning_Spatio-Temporal_Transformer_for_Visual_Tracking_ICCV_2021_paper.html)(ICCV2021)，结合带[MU](https://ieeexplore.ieee.org/document/9156764)的[SuperDimp](https://ieeexplore.ieee.org/document/9010649)算法，在线[MDNet](https://link.springer.com/chapter/10.1007/978-3-030-01225-0_6)分类器验证切换。**|     
17 |      | [**VOT2021**](https://www.votchallenge.net/vot2021/)  | RT | **TransT_M**(Multi-Template Transformer Tracking)：TransT_M is a variant of TransT [[8](https://ieeexplore.ieee.org/document/9578609/)]. We add a segmentation branch, a Multi-Template design, and an IoU prediction head on TransT, forming an end-to-end framework. We concatenate two templates in the spatial dimension, and input them into the template branch of TransT. IoU prediction head is a three-layer perceptron to predict the bounding box’s IoU and control the updating of the template. For details about TransT and the segmentation head, please refer to A.18.--From [《VOT2021 Report》](https://prints.vicos.si/publications/400)<br />**核心：[TransT算法](https://ieeexplore.ieee.org/document/9578609/)，代表性的Transformer Tracking算法，孪生匹配架构，CNN提取特征，Transformer机制重合模板和搜索区域，发表于CVPR2021。**|
18 |      | [**VOT2021**](https://www.votchallenge.net/vot2021/)  | ST | **RPTMask**：We propose a two-stage tracker, called RPTMask. The first stage is a base tracker responsible for locating the target bounding boxes. Specifically, we use ATOM [[16](https://ieeexplore.ieee.org/document/8953466)] to coarsely locate the target and update the tracking model and use RPT [[47](https://link.springer.com/chapter/10.1007/978-3-030-68238-5_43)] to generate the target bounding boxes. In the second stage, following STMVOS [[51](https://ieeexplore.ieee.org/document/9008790/)], we design a mask generation network to generate the target masks. First, only the first frame is set to be the memory frame. Second, we improve the space-time memory reader in STMVOS with the kernel trick [[56](https://link.springer.com/chapter/10.1007/978-3-030-58542-6_38)] and the top-k filtering [[10](https://ieeexplore.ieee.org/document/9577550)] strategy. Third, following AlphaRefine [[67](https://ieeexplore.ieee.org/document/9577854)], we add a refined box regression head paralleled to the mask decoder. The backbone in all models is ResNet50. --From [《VOT2021 Report》](https://prints.vicos.si/publications/400)<br />**核心：[ATOM](https://ieeexplore.ieee.org/document/8953466)算法粗定位，[RPT](https://link.springer.com/chapter/10.1007/978-3-030-68238-5_43)算法生成矩形框，[STMVOS](https://ieeexplore.ieee.org/document/9008790/)生成分割结果，[AlphaRefine](https://ieeexplore.ieee.org/document/9577854)再次修正结果。**|     
19 |      | [**VOT2020**](https://www.votchallenge.net/vot2020/)  | RGBD | **ATCAIS**(Accurate Tracking by Category-Agnostic Instance Segmentation for RGBD Image)：The proposed tracker combines both instance segmentation and the depth information for accurate tracking. ATCAIS is based on the ATOM tracker and the HTC instance segmentation method which is re-trained in a category-agnostic manner. The instance segmentation results are used to detect background distractors and to re-fine the target bounding boxes to prevent drifting. The depth value is used to detect the target occlusion or disappearance and re-finding the target. --From [《VOT2020 Report》](https://prints.vicos.si/publications/384)<br />**核心：[ATOM](https://ieeexplore.ieee.org/document/8953466)算法与[HTC](http://ieeexplore.ieee.org/document/8954166)实例分割相融合。**|
20 |      | [**VOT2020**](https://www.votchallenge.net/vot2020/)  | RGBT | **DFAT**(Decision Fusion Adaptive Tracker)：Decision Fusion Adaptive Tracker is based on Siamese structure. Firstly, the multi-layer deep features are extracted by Resnet-50. Then, multi-RPN module is employed to predict the central location with multi-layer deep features. Finally, an adaptive weight strategy for decision level fusion is utilized to generate the final result. In addition, the template features are updated by a linear template update strategy. --From [《VOT2020 Report》](https://prints.vicos.si/publications/384)<br />**核心：不同层次深度特征上应用RPN结构，再进行决策级融合。** |
21 |      | [**VOT2020**](https://www.votchallenge.net/vot2020/)  | LT | **LTMU_B**(A Baseline Long-Term Tracker with Meta-Updater)：The tracker LTMU_B is a simplified version of LTMU [[11](https://ieeexplore.ieee.org/document/9156764)] and LTDSE with comparable performance adding a RPN-based regression network, a sliding-window based re-detection module and a complex mechanism for updating models and target re-localization. The short-term tracker LTMU_B contains two components. One is for target localization and based on DiMP algorithm [[5](https://ieeexplore.ieee.org/document/9010649)] using ResNet50 as the backbone network. The update of DiMP is controlled by meta-updater which is proposed by LTMU (https://github.com/Daikenan/LTMU). The second component is the SiamMask network [[74](https://ieeexplore.ieee.org/document/8953931)] used for refining the bounding box after locating the centre of the target. It also takes the local search region as the input and outputs the tight bounding boxes of candidate proposals. For the verifier, we adopts MDNet network [[55](https://link.springer.com/chapter/10.1007/978-3-030-01225-0_6)] which uses VGGM as the backbone and is pre-trained on ILSVRC VID dataset. The classification score is finally obtained by sending the tracking result’s feature to three fully connected layers. GlobalTrack [[26](https://ojs.aaai.org/index.php/AAAI/article/view/6758)] is utilised as the global detector. --From [《VOT2020 Report》](https://prints.vicos.si/publications/384)<br />**核心：[LTMU算法](https://ieeexplore.ieee.org/document/9156764)，提出MU模块处理噪声样本更新问题，可适配于任何在线判决分类学习跟踪算法，发表于[CVPR2020](https://ieeexplore.ieee.org/document/9156764)。**|
22 |      | [**VOT2020**](https://www.votchallenge.net/vot2020/)  | RT | **AlphaRef**(Alpha-Refine)：We propose a simple yet powerful two-stage tracker, which consists of a robust base tracker (super-dimp) and an accurate refinement module named Alpha-Refine [[82](https://arxiv.org/abs/2007.02024)]. In the first stage, super-dimp robustly locates the target, generating an initial bounding box for the target. Then in the second stage, based on this result, Alpha-Refine crops a small search region to predict a high-quality mask for the tracked target. Alpha-Refine exploits pixel-wise correlation for fine feature aggregation, and uses non-local layer to capture global context informationBesides, Alpha-Refine also deploys a delicate mask prediction head [[60](https://ieeexplore.ieee.org/document/9156406)] to generate high-quality masks. The complete code and trained models of Alpha-Refine will be released at github.com/MasterBin-IIAU/AlphaRefine. --From [《VOT2020 Report》](https://prints.vicos.si/publications/384)<br />**核心：[Alpha-Refine算法](https://ieeexplore.ieee.org/document/9577854)，跟踪结果修正模块，类似缩小Search Region的SiamMask，用于两倍搜索区域，Pixel-wise卷积，Corner输出取代Box输出，Mask分支辅助。即插即用，可用于各类Tracker的结果优化，发表于CVPR2020。**|
23 |      | [**VOT2020**](https://www.votchallenge.net/vot2020/)  | ST | **RPT**(Learning Point Set Representation for Siamese Visual Tracking)：RPT tracker is formulated with a two-stage structure. The first stage is composed with two parallel subnets, one for target estimation with RepPoints [[84](https://ieeexplore.ieee.org/document/9009032)] in an offline-trained embedding space, the other trained online to provide high robustness against distractors [[13](https://ieeexplore.ieee.org/document/8953466)]. The online classification subnet is set to a lightweight 2-layer convolutional neural network. The target estimation head is constructed with Siamese-based feature extraction and matching. For the second stage, the set of RepPoints with highest confidence (i.e. online classification score) is fed into a modified D3S [[50](https://ieeexplore.ieee.org/document/9157386)] to obtain the segmentation mask. A segmentation map is obtained by combining enhanced target location channel with target and background similarity channels. The backbone is ResNet50 pretrained on ImageNet, while the target estimation head is trained using pairs of frames from YouTube-Bounding Box [[59](https://ieeexplore.ieee.org/document/8100272)], COCO [[45](https://link.springer.com/chapter/10.1007/978-3-319-10602-1_48)] and ImageNet VID [[63](https://link.springer.com/article/10.1007/s11263-015-0816-y)] datasets. --From [《VOT2020 Report》](https://prints.vicos.si/publications/384)<br />**核心：[RPT算法](https://link.springer.com/chapter/10.1007/978-3-030-68238-5_43)，预测点集而非目标框，综合了可形变卷积、RepPoints 检测、多层级卷积特征等思想，精度非常不错，但速度很慢。**|
24 |      | [**VOT2019**](https://www.votchallenge.net/vot2019/)  | RGBD | **SiamDW-D**(Online Deeper and Wider Siamese Networks for RGBD Visual Tracking)：SiamDW-D is a long-term tracker which mainly addresses the problems of target appearance variations, and frequent disappearance and re-appearance of target objects. It contains three parts, i.e. a main tracker, a re-detection module, and a multi-template matching module. The main tracker is based on [[109](https://ieeexplore.ieee.org/document/8953458)], and further equips it with an online updating model, similar to [[16](https://ieeexplore.ieee.org/document/8953466), [75](https://ieeexplore.ieee.org/document/7780834)]. The re-detection module is triggered when the main tracker is not confident on its predictions. However, in some cases, the confidence of main tracker is not reliable for re-detection module triggering, therefore we introduce a multi-template matching module. It matches the unreliable tracking results with history templates, and outputs a more reliable estimation. Moreover, the depth information is also used to estimate the disappearance of target objects. Code is available at https://github.com/researchmm/VOT2019. For more information see the short-term tracker from the same group (A.39). --From [《VOT2019 Report》](https://prints.vicos.si/publications/375)<br />**核心：[SiamDW算法](https://ieeexplore.ieee.org/document/8953458)，非常细致的Siamese Tracking讨论，发表于CVPR2019。**|
25 |      | [**VOT2019**](https://www.votchallenge.net/vot2019/)  | RGBT | **mfDiMP**(Multi-modal fusion for end-to-end RGB-T tracking)：The mfDiMP tracker contains an end-to-end tracking framework for fusing the RGB and TIR modalities in RGB-T tracking [[107](https://ieeexplore.ieee.org/document/9022263)]. The baseline tracker is DiMP (Discriminative Model Prediction) [[8](https://ieeexplore.ieee.org/document/9010649)], which employs a carefully designed target prediction network trained end-to-end using a discriminative loss. The mfDiMP tracker fuses modalities at the feature level on both the IoU predictor and the model predictor of DiMP [[107](https://ieeexplore.ieee.org/document/9022263)].--From [《VOT2019 Report》](https://prints.vicos.si/publications/375)<br />**核心：[Dimp算法](https://ieeexplore.ieee.org/document/9010649)结合多模态特征即为[mfDiMP算法](https://ieeexplore.ieee.org/document/902226)，上述两个工作分别发表于ICCV2019Main Conference和Workshop。**|
26 |      | [**VOT2019**](https://www.votchallenge.net/vot2019/)  | LT | **LT-DSE**(long-term tracking by diving videos into successive short episodes)：The tracker LT-DSE divides each long-term sequence into several short episodes and tracks the target in each episode using short-term tracking techniques. If the target disappears, the image-wide re-detection outputs the possible location and size of the target. The tracker crops the local search region and sends it to the RPN based regression network. Then, the candidate proposals from the regression network will be scored by the online learned verifier. The candidate with the maximum score will be regarded as the target and the tracker conducts short-term tracking which contains two components. One is for target localization based on ATOM algorithm [[16](https://ieeexplore.ieee.org/document/8953466)]. It uses ResNet18 as the backbone network and adds two convolutional layers above it. The other component is the SiamMask network [[96](https://ieeexplore.ieee.org/document/8953931)] used for refining the bounding box after locating the centre of the target. For the verifier RT-MDNet network [[42](https://link.springer.com/chapter/10.1007/978-3-030-01225-0_6)] is used as backbone and is pre-trained on ILSVRC VID dataset. The architecture of the region-proposal network is based on [[108](https://arxiv.org/abs/1809.04320)]. The network is trained using LaSOT dataset [[24](https://ieeexplore.ieee.org/document/8954084)] and ILSVRC image detection dataset. --From [《VOT2019 Report》](https://prints.vicos.si/publications/375)<br />**核心：[MBMD](https://arxiv.org/abs/1809.04320)算法全局重找回，[ATOM](https://ieeexplore.ieee.org/document/8953466)算法局部跟踪，[RT-MDNet](https://link.springer.com/chapter/10.1007/978-3-030-01225-0_6)验证跟踪结果判断切换，[SiamMask](https://ieeexplore.ieee.org/document/8953931)修正跟踪结果。比较工程化的算法，随重新梳理为[LTMU](https://ieeexplore.ieee.org/document/9156764)算法，发表于[CVPR2020](https://ieeexplore.ieee.org/document/9156764)。** |
27 |      | [**VOT2019**](https://www.votchallenge.net/vot2019/)  | RT | **SiamMargin**(Discriminative Siamese Embedding for Object Tracking)：SiamMargin is based on the SiamRPN++[58] algorithm and it learns discriminative embedding features in Siamese networks for object tracking. In the training stage, a discrimination loss is added to the embedding layer which imposed a margin to the decision boundary to encourage learning discriminative embeddings. The discriminative embedding is offline learned in the training phase, so it keeps the high speed of Siamese RPN. In the inference stage we exploit an online updating method with ROIAlign for Siamese networks based trackers. The template feature of the object in current frame is obtained by ROIAlign from features of the current search region. Then, the template feature is updated via a moving average strategy. The discriminative embedding features are leveraged to accommodate the appearance change with properly online updating. --From [《VOT2019 Report》](https://prints.vicos.si/publications/375)<br />**核心：[SiamRPN++算法](https://ieeexplore.ieee.org/document/8954116)。**|
28 |      | [**VOT2019**](https://www.votchallenge.net/vot2019/)  | ST | **ATP**(Accurate Tracking by Progressively refining) : We improve the tracking accuracy by progressively refining the target estimation. The process consists of three stages. At the first stage, we adopt the ATOM tracker [[16](https://ieeexplore.ieee.org/document/8953466)] to estimate the axis-aligned target position. At the second stage, we crop out an image patch centred on the estimated target and feed it into a segmentation network. Specifically, SiamMask [[96](https://ieeexplore.ieee.org/document/8953931)] is used. At the third stage, a rotated bounding box is generated from the segmentation. --From [《VOT2019 Report》](https://prints.vicos.si/publications/375)<br />**核心：[ATOM](https://ieeexplore.ieee.org/document/8953466)算法粗定位，[SiamMask](https://ieeexplore.ieee.org/document/8953931)算法精修正，上述两个算法均发表于CVPR2019。**|
29 |      | [**VOT2018**](https://www.votchallenge.net/vot2018/)  | LT | **MBMD**( MobileNet based tracking by detection algorithm)：The proposed tracker consists of a bounding box regression network and a verifier network. The regression network regresses the target object’s bounding box in a search region given the target in the first frame. Its outputs are several candidate boxes and each box’s reliability is evaluated by the verifier todetermine the predicted target box. If the predicted scores of both networks are below the thresholds, the tracker searches the target in the whole image. The regression network uses SSD-MobileNet architecture [[35](https://arxiv.org/abs/1704.04861),[52](https://link.springer.com/chapter/10.1007/978-3-319-46448-0_2)] and its parameters are fixed during online tracking. The verifier is similar to MDNet [[64](https://openaccess.thecvf.com/content_cvpr_2016/html/Nam_Learning_Multi-Domain_Convolutional_CVPR_2016_paper.html)] and is implemented by VGGM pretrained on ImageNet classification dataset. The last three layers’ parameters of the verifier are updated online to filter the distractors for the tracker. --From [《VOT2018 Report》](https://prints.vicos.si/publications/365)<br />**核心：[MBMD算法](https://arxiv.org/abs/1809.04320)，滑动窗搜索+类SiamRPN回归+MDNet验证，但速度慢。改进算法[SPLT](https://openaccess.thecvf.com/content_ICCV_2019/html/Yan_Skimming-Perusal_Tracking_A_Framework_for_Real-Time_and_Robust_Long-Term_Tracking_ICCV_2019_paper.html)精度类似，速度可达到实时，发表于[ICCV2019](https://openaccess.thecvf.com/content_ICCV_2019/html/Yan_Skimming-Perusal_Tracking_A_Framework_for_Real-Time_and_Robust_Long-Term_Tracking_ICCV_2019_paper.html)。** |
30 |      | [**VOT2018**](https://www.votchallenge.net/vot2018/)  | RT | **SiamRPN**(High Performance Visual Tracking with Siamese Region Proposal Network)：The tracker SiamRPN consists of a Siamese sub-network for feature extraction and a region proposal sub-network including the classification branch and regression branch. In the inference phase, the proposed framework is formulated as a local one-shot detection task. The template branch of the Siamese subnetwork is pre-computed while correlation layers are formulated as convolution layers to perform online tracking [[48](https://ieeexplore.ieee.org/document/8579033)]. What is more, SiamRPN introduces an effective sampling strategy to control the imbalanced sample distribution and make the model focus on the semantic distractors [[102](https://link.springer.com/chapter/10.1007/978-3-030-01240-3_7)]. --From [《VOT2018 Report》](https://prints.vicos.si/publications/365)<br />**核心：[SiamRPN跟踪框架](https://ieeexplore.ieee.org/document/8579033)，Siamese网络特征提取，互相关融合模板与搜索区域，RPN生成候选，分类回归头输出目标位置，发表于[CVPR2018](https://ieeexplore.ieee.org/document/8579033)。** |     
31 |      | [**VOT2018**](https://www.votchallenge.net/vot2018/)  | ST | **MFT**(Multi-solution Fusion for Visual Tracking)：MFT tracker is based on correlation filtering algorithm. Firstly, different multi-resolution features with continuous convolution operator [[15](https://ieeexplore.ieee.org/document/8100216)] are combined. Secondly, in order to improve the robustness a multi-solution using different features is trained and multi-solutions are optimally fused to predict the target location. Lastly, different combinations of Res50, SE-Res50, Hog, and CN features are applied to the different tracking situations. --From [《VOT2018 Report》](https://prints.vicos.si/publications/365)<br />**核心：相关滤波结合多种深浅层特征。** |
32 |      | [**VOT2017**](https://www.votchallenge.net/vot2017/)  | TIR | **DSLT**(Dense Structural Learning based Tracker)：DSLT extends Struck with the ability to learn from dense samples and high dimensional features. DSLT runs in a simple tracking by detection mode, with no scale adaptation and occlusion handling. Compared to our initial work [[79](https://www.sciencedirect.com/science/article/abs/pii/S0167865517301654)], the feature representation in DSLT consists of 28D HOG features computed without block normalization as well as a 1D motion feature which is simply the absolute difference between consecutive frames. The search range is also set larger to account for faster motion and to get more training samples. --From [《VOT2017 Report》](https://openaccess.thecvf.com/content_ICCV_2017_workshops/papers/w28/Kristan_The_Visual_Object_ICCV_2017_paper.pdf)<br />**核心：[DSLT算法](https://www.sciencedirect.com/science/article/abs/pii/S0167865517301654)，密集采样，Structured SVM和相关滤波相结合，发表于[PRL2018](https://www.sciencedirect.com/science/article/abs/pii/S0167865517301654)。** | 
33 |      | [**VOT2017**](https://www.votchallenge.net/vot2017/)  | RT | **SiamFC**(Fully-Convolutional Siamese Network)：SiamFC [[5](https://link.springer.com/chapter/10.1007/978-3-319-48881-3_56)] applies a fully-convolutional Siamese network trained to locate an exemplar image within a larger search image. The network is fully convolutional with respect to the search image: dense and efficient slidingwindow evaluation is achieved with a bilinear layer that computes the cross-correlation of two inputs. The deep conv-net is trained offline on the ILSVRC VID dataset [[56](https://link.springer.com/article/10.1007/s11263-015-0816-y)] to address a general similarity learning problem. This similarity function is then used within a simplistic tracking algorithm. The architecture of the conv-net resembles ‘AlexNet’ [[34](https://papers.nips.cc/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html)]. This version of SiamFC incorporates some minor improvements and is available as the baseline model of the CFNet paper [[64](https://openaccess.thecvf.com/content_cvpr_2017/html/Valmadre_End-To-End_Representation_Learning_CVPR_2017_paper.html)]. --From [《VOT2017 Report》](https://openaccess.thecvf.com/content_ICCV_2017_workshops/papers/w28/Kristan_The_Visual_Object_ICCV_2017_paper.pdf)<br />**核心：[SiamFC算法](https://link.springer.com/chapter/10.1007/978-3-319-48881-3_56)，Siamese Tracking鼻祖，发表于[ECCVW2016](https://link.springer.com/chapter/10.1007/978-3-319-48881-3_56)。** |   
34 |      | [**VOT2017**](https://www.votchallenge.net/vot2017/)  | ST | **CFCF**(Convolutional Features for Correlation Filters)：The tracker ‘CFCF’ is based on the feature learning study in [[23](https://ieeexplore.ieee.org/abstract/document/8291524)] and the correlation filter based tracker in [[17](https://link.springer.com/chapter/10.1007/978-3-319-46454-1_29)]. The proposed tracker employs a fully convolutional neural network (CNN) model trained on ILSVRC15 [[56](https://link.springer.com/article/10.1007/s11263-015-0816-y)] video dataset by the introduced learning framework in [[23](https://ieeexplore.ieee.org/abstract/document/8291524)]. This framework is designed for correlation filter formulation in [[13](http://www.bmva.org/bmvc/2014/papers/paper038/index.html)]. To learn features, convolutional layers of VGGM-2048 network [[10](http://www.bmva.org/bmvc/2014/papers/paper054/index.html)], which is trained on [18], with an extra convolutional layer is fine-tuned on ILSVRC15 dataset. The first, fifth and sixth convolutional layers of the learned network, HOG [[47](https://ieeexplore.ieee.org/document/1467360)] and Colour Names (CN) [[65](https://ieeexplore.ieee.org/document/4982667)] are integrated to the tracker of [[17](https://link.springer.com/chapter/10.1007/978-3-319-46454-1_29)]. The reader is referred to [[23](https://ieeexplore.ieee.org/abstract/document/8291524)] for details. --From [《VOT2017 Report》](https://openaccess.thecvf.com/content_ICCV_2017_workshops/papers/w28/Kristan_The_Visual_Object_ICCV_2017_paper.pdf)<br />**核心：相关滤波结合浅深层特征。** |   
35 |      | [**VOT2016**](https://www.votchallenge.net/vot2016/)  | TIR | **EBT**(Edge Box Tracker)：EBT tracker is not limited to a local search window and has ability to probe efficiently the entire frame. It generates a small number of ‘high-quality’ proposals by a novel instance-specific objectness measure and evaluates them against the object model that can be adopted from an existing tracking-by-detection approach as a core tracker. During the tracking process, it updates the object model concentrating on hard false-positives supplied by the proposals, which help suppressing distractors caused by difficult background clutters, and learns how to re-rank proposals according to the object model. Since the number of hypotheses the core tracker evaluates is reduced significantly, richer object descriptors and stronger detectors can be used. More details can be found in [[25](https://ieeexplore.ieee.org/document/7780477)]. --From [《VOT-TIR2016 Report》](https://data.votchallenge.net/vot2016/presentations/vot_tir_2016_paper.pdf)<br />**核心：[EBT算法](https://ieeexplore.ieee.org/document/7780477)，Edge Box提取全局Proposal，在线更新Structured SVM分类，发表于[CVPR2016](https://ieeexplore.ieee.org/document/7780477)。**  |  
36 |      | [**VOT2016**](https://www.votchallenge.net/vot2016/)  | ST | **TCNN**(Tree-structured Convolutional Neural Network Tracker)：TCNN [[96](https://arxiv.org/abs/1608.07242)] maintains multiple target appearance models based on CNNs in a tree structure to preserve model consistency and handle appearance multi-modality effectively. TCNN tracker consists of two main components, state estimation and model update. When a new frame is given, candidate samples around the target state estimated in the previous frame are drawn, and the likelihood of each sample based on the weighted average of the scores from multiple CNNs is computed. The weight of each CNN is determined by the reliability of the path along which the CNN has been updated in the tree structure. The target state in the current frame is estimated by finding the candidate with the maximum likelihood. After tracking a predefined number of frames, a new CNN is derived from an existing one, which has the highest weight among the contributing CNNs to target state estimation.-From [《VOT2016 Report》](https://data.votchallenge.net/vot2016/presentations/vot_2016_paper.pdf)<br />**核心：[TCNN算法](https://arxiv.org/abs/1608.07242)，稀疏候选采样，树结构集成CNN分类及在线更新。**  | 
37 |      | [**VOT2015**](https://www.votchallenge.net/vot2015/)  | TIR | **sPST**(simplified Proposal Selection Tracker)：The simplified Proposal Selection Tracker (sPST) is based on our ICCV2015 paper [[21](https://ieeexplore.ieee.org/document/7410711)]. sPST operates in two phases. Firstly, we propose a set of candidate object locations computed by tracking-by-detection framework [[35](https://ieeexplore.ieee.org/document/6619152/)], where we use the frame as is and rotate them according to the ground truth annotation in the initial frame if applicable. Secondly, we determine the best candidate as the tracking result by two cues: detection confidence score and an objectness measure computed with edges [[48](https://link.springer.com/chapter/10.1007/978-3-319-10602-1_26)]. Note that the full version of our tracker uses additional proposals and motion boundaries calculated with optical flow. But it is not included in this submission due to the computational cost of the optical flow method. The reader is referred to [[21](https://ieeexplore.ieee.org/document/7410711)] for details.--From [《VOT-TIR2015 Report》](https://data.votchallenge.net/vot2015/presentations/vot_tir_2015_paper.pdf)<br />**核心：[PST算法](https://ieeexplore.ieee.org/document/7410711)，先生成Proposal再利用SVM检测置信值+Objectness来选择，发表于[ICCV2015](https://ieeexplore.ieee.org/document/7410711)。**|
38 |      | [**VOT2015**](https://www.votchallenge.net/vot2015/)  | ST | **MDNet**(Multi-Domain Convolutional Neural Network Tracker)：MDNet tracker represents the target object using a Convolutional Neural Network (CNN). MDNet pre-trains the CNN using a set of videos with tracking ground-truth annotations to obtain a generic representation for an arbitrary new sequence. The network is composed of two partsshared layers and domain specific layers, where domains correspond to individual tracking sequences and each domain has a separate branch for binary classification. After training, a generic representation in the shared layers across all domains is obtained. The tracking is performed by sampling target candidates around the previous target state, evaluating them on the CNN, and identifying the sample with the maximum score. For more details, the interested reader is referred to [[52](https://arxiv.org/pdf/1510.07945.pdf)].--From [《VOT2015 Report》](https://data.votchallenge.net/vot2015/presentations/vot_2015_paper.pdf)<br />**核心：[MDNet算法](https://openaccess.thecvf.com/content_cvpr_2016/html/Nam_Learning_Multi-Domain_Convolutional_CVPR_2016_paper.html)，早期经典深度学习跟踪算法，利用稀疏采样+深度分类，发表于[CVPR2016](https://openaccess.thecvf.com/content_cvpr_2016/html/Nam_Learning_Multi-Domain_Convolutional_CVPR_2016_paper.html)。**|
39 |      | [**VOT2014**](https://www.votchallenge.net/vot2014/)  | ST | **DSST**(Discriminative Scale Space Tracker)：The Discriminative Scale Space Tracker (DSST) [[14](http://www.bmva.org/bmvc/2014/papers/paper038/index.html)]  extends the Minimum Output Sum of Squared Errors (MOSSE) tracker [[5](https://ieeexplore.ieee.org/document/5539960/)] with robust scale estimation. The MOSSE tracker works by training a discriminative correlation filter on a set of observed sample grey scale patches. This correlation filter is then applied to estimate the target translation in the next frame. The DSST additionally learns a one-dimensional discriminative scale filter, that is used to estimate the target size. For the translation filter, the intensity features employed in the MOSSE tracker is combined with a pixel-dense representation of HOG-features.  --From [《VOT2014 Report》](https://www.votchallenge.net/vot2014/download/vot_2014_paper.pdf)<br /> **核心：相关滤波算法[DSST](http://www.bmva.org/bmvc/2014/papers/paper038/index.html)，发表于[BMVC2014](http://www.bmva.org/bmvc/2014/papers/paper038/index.html)。**|
40 |      | [**VOT2013**](https://www.votchallenge.net/vot2013/)  | ST | **PLT**(Single scale pixel based LUT tracker)：PLT runs a classifier at a fixed single scale for each test image, to determine the top scoring bounding box which is then the result of object detection. The classifier uses a binary feature vector constructed from color, grayscale and gradient information. To select a small set of discriminative features, an online sparse structural SVM [[20](https://ieeexplore.ieee.org/document/6126251)] is used. Since the object can be non-rigid and the bounding box may be noisy, not all pixels in the bounding box belong to the object. Hence, a probabilistic object-background segmentation mask from color histograms is created and used to weight the features during SVM training. The resulting weighted and convex problem can be solved in three steps: (i) compute the probability that a pixel belongs to the objectby using its color. (ii) solve the original non-sparse structural SVM and (iii) shrink the solution [[21](https://ieeexplore.ieee.org/document/6248061)], i.e. features with smallest values are discarded. Since the feature vector is binary, the linear classifier can be implemented as a lookup table for fast speed.  --From [《VOT2013 Report》](https://www.votchallenge.net/vot2013/Download/vot_2013_paper.pdf) <br /> **核心：在线结构SVM分类器，[Struck](https://ieeexplore.ieee.org/document/6126251)跟踪算法，发表于[CVPR2011](https://ieeexplore.ieee.org/document/6126251)。**|
41 | 


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