Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
TGFormer: Transformer with Track Query Group for Multi-Object Tracking
Authors: Rui Zeng, Yuanzhou Huang, Songwei Pei
AAAI 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that our method achieves competitive performance on the MOT Challenge and Dance Track datasets. Extensive ablation experiments further demonstrate the effectiveness of our method. |
| Researcher Affiliation | Academia | School of Computer Science (National Pilot Software Engineering School), Beijing University of Posts and Telecommunications, Beijing 100876, China {zengrui@, huangyuanzhou@, peisongwei@}bupt.edu.cn |
| Pseudocode | No | The paper describes the method using prose and mathematical equations but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any explicit statement about releasing source code for TGFormer, nor does it include links to a code repository. |
| Open Datasets | Yes | We conduct experiments on the MOT Challenge (Milan et al. 2016; Dendorfer et al. 2020) and Dance Track datasets (Sun et al. 2022). The MOT17 benchmark (Milan et al. 2016) includes 7 training sequences and 7 testing sequences... For this combined dataset, we train for 130 epochs... We also incorporate the Crowd Human (Shao et al. 2018) Validation set. |
| Dataset Splits | Yes | The MOT17 benchmark (Milan et al. 2016) includes 7 training sequences and 7 testing sequences... The MOT20 benchmark (Dendorfer et al. 2020) has 4 training and 4 testing sequences... For all ablation experiments, we split the sequences in the MOT17 training set into two halves, using one half as the training set and the other half as the validation set. |
| Hardware Specification | Yes | Training uses 4 NVIDIA A800 GPUs with a batch size of 1 per GPU, each batch containing a video clip with multiple frames. |
| Software Dependencies | No | TGFormer builds on Me MOTR (Gao and Wang 2023) with Res Net50 as the backbone and DAB-Deformable DETR pretrained on COCO as the detector. Training uses... The Adam W optimizer with a 2.0 10 4 learning rate is applied. This text lists software components but lacks specific version numbers for reproducibility. |
| Experiment Setup | Yes | Training uses 4 NVIDIA A800 GPUs with a batch size of 1 per GPU, each batch containing a video clip with multiple frames. The Adam W optimizer with a 2.0 10 4 learning rate is applied. Targets with scores below ̷̲̣̔̄̂͂̀̂̀͂̓ = 0.5 or Io U below ̷̲̣̔̄̂͂̀̂̀͂̓̂̂̃ = 0.5 are filtered... The confidence thresholds are set as ̷̲̣̔̄̂͂̀̂̀͂̓̈̇̂̂̀̂ = 0.85, ̷̲̣̔̄̂͂̀̂̀͂̓̂̀̂ = 0.7, ̷̲̣̔̄̂͂̀̂̀͂̓̅̃̇ = 0.5... For this combined dataset, we train for 130 epochs, reducing the learning rate tenfold at the 120th epoch. The number of clip frames increases from the original 2 frames to 3, 4, 5, and 6 frames at the 50th, 70th, 90th, and 120th epochs, respectively. |