Online Decision Based Visual Tracking via Reinforcement Learning

Authors: ke Song, Wei Zhang, Ran Song, Yibin Li

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive results show that our DTNet achieves stateof-the-art tracking performance as well as a good balance between accuracy and efficiency.
Researcher Affiliation Academia Ke Song Wei Zhang Ran Song Yibin Li School of Control Science and Engineering, Shandong University, Jinan, China songke_vsislab@mail.sdu.edu.cn {davidzhang,ransong,liyb}@sdu.edu.cn
Pseudocode Yes Please refer to Algorithm 1 in the supplementary material available at the website mentioned in the abstract for the details of the whole training process.
Open Source Code No The project website is available at https://vsislab.github. io/DTNet/. The paper provides a link to a 'project website' but does not explicitly state that source code for the methodology is available at this link or elsewhere.
Open Datasets Yes In this section, we conduct comparative evaluations on the benchmarks including OTB-2013 [37], OTB-50 [38], OTB-100 [38], La SOT [12], Tracking Net [24], UAV123 [23] and VOT18 [18]... The sequences from VID [28] and Youtube_BB [27] datasets are used to train the DTNet
Dataset Splits No The sequences from VID [28] and Youtube_BB [27] datasets are used to train the DTNet including the decision and the tracker modules for 6 × 10^5 episodes with Adam optimizer. The paper mentions datasets used for training and evaluation on benchmarks, but does not explicitly state a validation split percentage, sample counts, or a specific methodology for creating a validation set within its own experimental setup.
Hardware Specification Yes The experiments were implemented in Py Torch on a computer with a 3.70GHz Intel Core i7-8700K CPU and two NVIDIA GTX 1080Ti GPUs.
Software Dependencies No The experiments were implemented in Py Torch on a computer with a 3.70GHz Intel Core i7-8700K CPU and two NVIDIA GTX 1080Ti GPUs. The paper mentions 'Py Torch' but does not specify a version number or other software dependencies with their versions.
Experiment Setup Yes We set the capacity of the replay buffer to 5000, the learning rate to 0.0001, the discount factor γ in Equ. 1 to 0.2, the batch size to 128 and nκ is set to 3 × 10^5. For the ϵ-greedy algorithm, ϵ is set to 1 and decays to 0.1 gradually.