Sparse Graph Tracker (SGT)

Official code for Sparse Graph Tracker (SGT) based on the Detectron2 framework. Please feel free to leave an ISSUE or send me an email ([email protected]).

News

• (2022.10.11) Our paper is accepted WACV 2023! (arxiv paper will be updated soon)
• (2022.10.06) Code and pretrained weights are released!

Installation

• Please refer INSTALL.md for the details

Dataset Setup

• Please refer DATASET.md for the details

Model Zoo

• download model weights

• Please modify the path of checkpoints in the config file based on your checkpoint directory

MOT17

Name	Dataset	HOTA	MOTA	IDF1	Download
SGT	MOT17	58.2	73.2	70.2	model
SGT	MOT17 + CrowdHuman	60.8	76.4	72.8	model

MOT20

Name	Dataset	HOTA	MOTA	IDF1	Download
SGT	MOT20	51.6	64.5	62.7	model
SGT	MOT20 + CrowdHuman	57.0	72.8	70.6	model

HiEve

Name	Dataset	MOTA	IDF1	Download
SGT	HiEve	47.2	53.7	model

How to run?

Train

python projects/SGT/train_net.py --config-file projects/SGT/configs/MOT17/sgt_dla34.yaml --data-dir /root/datasets --num-gpus 2 OUTPUT_DIR /root/sgt_output/mot17_val/dla34_mot17-CH

Inference

python projects/SGT/train_net.py --config-file projects/SGT/configs/MOT17/sgt_dla34.yaml --data-dir /root/datasets --num-gpus 1 --eval-only OUTPUT_DIR /root/sgt_output/mot17_test/dla34_mot17-CH

Visualization

## GT
python projects/Datasets/MOT/vis/vis_gt.py --data-root <$DATA_ROOT> --register-data-name <e.g., mot17_train> 
python projects/Datasets/MOT/vis/vis_gt.py --data-root <$DATA_ROOT> --register-data-name <e.g., mix_crowdhuman_train> --no-video-flag 


## model output
python projects/Datasets/MOT/vis/vis_seq_from_txt_result.py --data-root <$DATA_ROOT> --result-dir <$OUTPUT_DIR> --data-name {mot17, mot20, hieve, mot17_sub, mot20_sub} --tgt-split {val,test}

Motivation

Pipeline

MOT Benchmark Results

Ablation Experiment Results

Visualization

License

Code of SGT is licensed under the CC-BY-NC 4.0 license and free for research and academic purpose. SGT is based on the framework Detectron2 which is released under the Apache 2.0 license and the detector CenterNet which is released under the MIT license. This codebase also provides Detectron2 version of FairMOT which is released under the MIT license.

Citation

@inproceedings{hyun2023detection,
  title={Detection recovery in online multi-object tracking with sparse graph tracker},
  author={Hyun, Jeongseok and Kang, Myunggu and Wee, Dongyoon and Yeung, Dit-Yan},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={4850--4859},
  year={2023}
}