ZoomTrack: Target-aware Non-uniform Resizing for Efficient Visual Tracking
Authors: Yutong Kou, Jin Gao, Bing Li, Gang Wang, Weiming Hu, Yizheng Wang, Liang Li
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments on five challenging datasets based on two kinds of transformer trackers, i.e., OSTrack and Trans T, demonstrate consistent improvements over them. |
| Researcher Affiliation | Collaboration | Yutong Kou1,2, Jin Gao1,2 , Bing Li1,5, Gang Wang4 , Weiming Hu1,2,3, Yizheng Wang4, Liang Li4 1State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS), CASIA 2School of Artificial Intelligence, University of Chinese Academy of Sciences 3School of Information Science and Technology, Shanghai Tech University 4Beijing Institute of Basic Medical Sciences 5People AI, Inc |
| Pseudocode | No | The paper describes the approach using mathematical formulations and descriptive text, but does not include a dedicated pseudocode or algorithm block. |
| Open Source Code | Yes | Codes and models are available at https://github.com/Kou-99/Zoom Track. |
| Open Datasets | Yes | Comprehensive experiments are conducted on five large-scale benchmarks, including GOT-10k [16], La SOT [11], La SOText [12], TNL2k [30], Tracking Net [23]. |
| Dataset Splits | No | The paper mentions 'training split' and 'test splits' for datasets like GOT-10k, but does not explicitly provide detailed information about a separate 'validation' split in terms of percentages or counts, nor does it explicitly state the use of a standard validation split. |
| Hardware Specification | Yes | The trackers are trained on four NVIDIA V100 GPUs. The inference speed and MACs are evaluated on a platform with Intel i7-11700 CPU and NVIDIA V100 GPU. |
| Software Dependencies | Yes | This convex objective function can be efficiently solved by a standard QP solver [20] (see Appendix A for more detailed derivations). [20] Martin S. Andersen, Joachim Dahl, and Lieven Vandenberghe. CVXOPT: A python package for convex optimization (version: 1.3.0), 2022. |
| Experiment Setup | Yes | The search patch size (256 256) is the same as the baseline methods, whereas the context factor is enlarged from 4 to 5. We use a controllable grid g with shape 17 17(m = n = 16) . The importance score map is generated by a Gaussian function with bandwidth β = 64. We set magnification factor γ = 1.5 and λ = 1 in Eq. (9). We use smaller jitter js = 0.1 with probability 0.8 and larger jitter jl = 0.5 with probability 0.2. |