Home

Awesome

Sparse Graph Tracker (SGT)

PWC PWC PWC PWC

Official code for Sparse Graph Tracker (SGT) based on the Detectron2 framework. Please feel free to leave an ISSUE or send me an email (jhyunaa@ust.hk).

News

Installation

Dataset Setup

Model Zoo

MOT17

NameDatasetHOTAMOTAIDF1Download
SGTMOT1758.273.270.2model
SGTMOT17 + CrowdHuman60.876.472.8model

MOT20

NameDatasetHOTAMOTAIDF1Download
SGTMOT2051.664.562.7model
SGTMOT20 + CrowdHuman57.072.870.6model

HiEve

NameDatasetMOTAIDF1Download
SGTHiEve47.253.7model

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

image

Pipeline

image

MOT Benchmark Results

image

Ablation Experiment Results

image

image

Visualization

image

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}
}