ColTrack

ColTrack not only outperforms state-of-the-art methods on large-scale datasets under high frame rates but also achieves higher and more stable performance under low frame rates. This allows it to obtain a higher equivalent FPS by reducing the frame rate requirement.

ColTrack: Collaborative Tracking Learning for Frame-Rate-Insensitive Multi-Object Tracking

Yiheng Liu, Junta Wu, Yi Fu

arXiv:2308.05911

News

• (2023.07) Our paper is accepted by ICCV 2023!

Tracking performance

Dataset	HOTA	MOTA	IDF1
MOT17	61.0	78.8	73.9
Dancetrack	72.6	92.1	74.0
Dancetrack(+val)	75.3	92.2	77.3

Visualization results on MOT challenge test set

Installation

The codebase is built on top of DINO and MOTR.

Requirements

• Install pytorch using conda (optional)

  conda create -n coltrack python=3.9
  conda activate coltrack
  conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch

• Other requirements

  pip install Cython
  pip install -r requirements.txt

• Compiling CUDA operators

  cd models/dino/ops
   python setup.py build install
   # unit test (should see all checking is True)
   python test.py
   cd ../../..

Data preparation

Download MOT17, CrowdHuman, DanceTrack and unzip them under Coltrack_HOME as follows:

mot_files
   |——————dataset
   |        └——————dancetrack
   |                  └—————test
   |                  └—————train
   |                  └—————val
   |        └——————MOT17
   |                  └—————images
   |                          └—————train
   |                          └—————test
   └——————models
   |         └——————coltrack
   |                    └——————dancetrack_val.pth
   |         └——————dino
   |         └——————dino_e2e

Model zoo

Ablation model

• Standard models

  Put these models in the directory mot_files/models/coltrack.

  Model	HOTA	MOTA	IDF1
  dancetrack_val	73.51	92.1	76.48

• Dependency models

  These models are downloaded from DINO

  Model	Target folder
  checkpoint0027_5scale_swin	mot_files/models/dino
  checkpoint0029_4scale_swin	mot_files/models/dino
  checkpoint0031_5scale	mot_files/models/dino
  checkpoint0033_4scale	mot_files/models/dino

Training

Pretraining (Baseline model): This model is used to initialize the end-to-end model for the next stage.

cd <Coltrack_HOME>
bash scripts/train_dancetrack_val_tbd.sh your_log_folder_name

E2E model training: Select one model from the previous stage and put it in mot_files/models/dino_e2e/4scale_ablation_res_dancetrack.pth.

cd <Coltrack_HOME>
bash scripts/train_dancetrack_val_coltrack.sh your_log_folder_name

Tracking

cd <Coltrack_HOME>
bash scripts/test_dancetrack_val_coltrack.sh your_log_folder_name

# logs/your_log_folder_name/ColTrack/track_results: The tracking results of ColTrack.
# logs/your_log_folder_name/IPTrack/track_results : interpolated tracking results of ColTrack, which may be better or worse. The interpolation algorithm is motlib/tracker/interpolation/gsi.py

Demo for videos

cd <Coltrack_HOME>
# your_videos_path
#    |——————videos1.mp4
#    |——————videos2.mp4
bash scripts/demo.sh --output_dir your_log_folder_name --infer_data_path your_videos_path --is_mp4 --draw_tracking_results --inference_sampler_interval 1 --resume mot_files/models/coltrack/dancetrack_val.pth --options config_file=config/E2E/coltrack_inference.py

--inference_sampler_interval 3 : The downsampling interval of the video frames.

Demo for video frames

cd <Coltrack_HOME>

# your_video_frames_path
#    |——————videos1
#    |         └—————frame1.xxx
#    |         └—————frame2.xxx
#    |——————videos2
#    |         └—————xxx.xxx
#    |         └—————xxx.xxx

bash scripts/demo.sh --output_dir your_log_folder_name --infer_data_path your_video_frames_path --draw_tracking_results --inference_sampler_interval 1 --resume mot_files/models/coltrack/dancetrack_val.pth --options config_file=config/E2E/coltrack_inference.py

Citation

@article{liu2023collaborative,
  title={Collaborative Tracking Learning for Frame-Rate-Insensitive Multi-Object Tracking},
  author={Liu, Yiheng and Wu, Junta and Fu, Yi},
  booktitle={ICCV},
  year={2023}
}

Acknowledgement

A large part of the code is borrowed from DINO, MOTR, ByteTrack and Bot-SORT. Many thanks for their wonderful works.