One More Check: Making “Fake Background” Be Tracked Again

Authors: Chao Liang, Zhipeng Zhang, Xue Zhou, Bing Li, Weiming Hu1546-1554

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the proposed OMC on three MOT Challenge 1 benchmarks: MOT16 (Milan et al. 2016), MOT17 (Milan et al. 2016) and MOT20 (Dendorfer et al. 2020). Our method achieves new state-of-the-art MOTA and IDF1 on all three benchmarks. Furthermore, compared with other trackers using temporal cues under the same public detection protocol (Milan et al. 2016), our method still achieves better tracking performance on MOTA and IDF1. The experimental results demonstrate that our tracker OMC not only outperforms CSTrack largely, but also achieves new state-of-the-art MOTA and IDF1 scores on all three benchmarks.
Researcher Affiliation Academia 1School of Automation Engineering, University of Electronic Science and Technology of China (UESTC) 2 NLPR, Institute of Automation, Chinese Academy of Sciences (CASIA) 3 Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China (UESTC) 4 Intelligent Terminal Key Laboratory of Si Chuan Province
Pseudocode Yes Algorithm 1: Modified Cross-correlation, Py Torch-like def D (Eid t 1, F id t ): E = torch.Tensor(Eid t 1).view(n, c) # convert list to matrix F = F id t .view(c, h w) # reshape tensor to matrix M = torch.matmul(E, F) # matrix multiplication return M.view(h, w, n) # reshape matrix to tensor
Open Source Code Yes Code is released at https://github.com/JudasDie/SOTS.
Open Datasets Yes We evaluate the proposed OMC on three MOT Challenge 1 benchmarks: MOT16 (Milan et al. 2016), MOT17 (Milan et al. 2016) and MOT20 (Dendorfer et al. 2020). Following the common practices in MOT Challenge (Milan et al. 2016), we employ the CLEAR metric (Bernardin and Stiefelhagen 2008), particularly MOTA (the primary metric of MOT) and IDF1 (Ristani et al. 2016) to evaluate the overall performance. We use the same training data as CSTrack, including ETH (Ess et al. 2008), City Person (Zhang, Benenson, and Schiele 2017), Cal Tech (Doll ar et al. 2009), MOT17 (Milan et al. 2016), CUDK-SYSU (Xiao et al. 2017), PRW (Zheng et al. 2017) and Crowd Human (Shao et al. 2018).
Dataset Splits Yes In the second stage, we train the proposed recheck network while fixing the basic tracker s parameters on MOT17 (Milan et al. 2016) training set.
Hardware Specification Yes Our tracker is implemented using Python 3.7 and Py Torch 1.6.0. The experiments are conducted on a single RTX 2080Ti GPU and Xeon Gold 5218 2.30GHz CPU.
Software Dependencies Yes Our tracker is implemented using Python 3.7 and Py Torch 1.6.0.
Experiment Setup Yes The training procedure consists of two stages, i.e., basic tracker training and re-check network optimization. In the first stage... The network is trained with a SGD optimizer for 30 epochs. The batch size is 8. The initial learning rate is 5 10 4, and it decays to 5 10 5 at the 20th epoch... We set r in Eq. 5 to 3 and initialize the scale parameter h in Eq. 12 to 10.