POST: POlicy-Based Switch Tracking

Authors: Ning Wang, Wengang Zhou, Guojun Qi, Houqiang Li12184-12191

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive ablation studies and experimental comparisons against state-of-the-art trackers on 5 prevalent benchmarks verify the effectiveness of the proposed method. In this section, we first introduce the experimental details. Then, we analyze the effectiveness of our policy-based ensemble framework. Finally, we compare our method with state-of-the-art trackers on 5 standard tracking benchmarks including OTB-2013 (Wu, Lim, and Yang 2013), OTB-2015 (Wu, Lim, and Yang 2015), Temple-Color (Liang, Blasch, and Ling 2015), UAV123 (Mueller, Smith, and Ghanem 2016), and La SOT (Fan et al. 2019).
Researcher Affiliation Collaboration Ning Wang,1 Wengang Zhou,1 Guojun Qi,2 Houqiang Li1 1CAS Key Laboratory of GIPAS, University of Science and Technology of China 2Futurewei Technologies wn6149@mail.ustc.edu.cn, guojun.qi@huawei.com, {zhwg, lihq}@ustc.edu.cn
Pseudocode Yes Algorithm 1: Training of Deep Q-Network
Open Source Code No The paper does not contain any explicit statement about making the source code available or provide a link to a code repository.
Open Datasets Yes 5 standard tracking benchmarks including OTB-2013 (Wu, Lim, and Yang 2013), OTB-2015 (Wu, Lim, and Yang 2015), Temple-Color (Liang, Blasch, and Ling 2015), UAV123 (Mueller, Smith, and Ghanem 2016), and La SOT (Fan et al. 2019).
Dataset Splits No The paper mentions using videos from VOT-2013, VOT-2014, and VOT-2015 for training the agent and La SOT's provided training videos. However, it does not explicitly provide specific training/validation/test dataset splits (e.g., percentages or sample counts) for reproducibility.
Hardware Specification Yes The experimental environment is Py Torch on a computer with 4.00GHz Intel Core I7-4790K and NVIDIA GTX 1080Ti GPU.
Software Dependencies No The paper mentions 'Py Torch' as the experimental environment but does not specify its version number or any other software dependencies with version numbers.
Experiment Setup Yes We randomly choose continuous 20 to 40 frames in a video for each episode, and our agent is trained for 2 105 episodes using Adam optimizer. The size of the replay buffer M is set to 10000. The learning rate is 0.0001, discount factor γ = 0.9 and batch size is 128. As for the ϵgreedy, we initially set ϵ = 1 and reduce it by 5% every 2000 episodes and fix it after it is reduced to 0.1.