Self-Supervised Multi-Object Tracking with Path Consistency

We propose a novel concept of path consistency to learn robust object matching without using manual object identity supervision. Our key idea is that, to track a object through frames, we can obtain multiple different association results from a model by varying the frames it can observe, i.e., skipping frames in observation. As the differences in observations do not alter the identities of objects, the obtained association results should be consistent. Based on this rationale, we formulate new Path Consistency Loss and have achieved SOTA on three tracking datasets (MOT17, PersonPath22, KITTI).

Preparation

Install dependencies

pip3 install -r requirements.txt
cd libs
git clone https://github.com/ifzhang/ByteTrack.git
git clone https://github.com/JonathonLuiten/TrackEval.git

Preprocess data

MOT17

• download MOT17 data, unzip it and place it as data/mot17
• download Tracktor++ detection from drive and place it as data/mot17/tracktor_det
• preprocess MOT17 data via python -m src.script.preprocesss_mot17

PersonPath

• download PersonPath22 dataset.
• use src/script/extract_pp_frames.py to extract frames and place them in data/personpath22/frames.
• preprocess PersonPath22 data via python -m src.script.preprocess_personpath.py.

Train model

# train MOT17
python -m src.train --cfg src/configs/mot.yaml --set aux.gpu $GPUID aux.runid 1
# train PersonPath22
python -m src.train --cfg src/configs/personpath_fcos.yaml --set aux.gpu $GPUID aux.runid 1 

For example, the first command will train model on MOT17 public detections. Following prior works, the detections are refined by Tracktor++. Results will be saved under log, where ckpts folder stores model weights. saves folder stores the accuracy of learned matching probability. Losses and accuracies are also visualized via wandb.

Evaluate model

# eval MOT17
python -m src.inference_mot --gpu $GPUID --exp log/mot17/public/all/mot/1/
# eval PersonPath22
python -m src.inference_personpath --gpu $GPUID --exp log/personpath22/fcos_processed/all/personpath_fcos-replicate/1/ 

The first command will evaluate model on MOT17 videos, save results to output/MOT17, and print the accuracy for training videos. Accuracy for test videos can be obtained by submitting to MOT17 server.

The second command will evaluate model on PersonPath22 test videos.

Pretrained Model

Pretrained model and results for MOT17 and PersonPath22 can be downloaded from drive. Each zip file contains log and output folders. The performance are:

In Progress

The training configs for PersonPath and KITTI are exactly the same as defined in src/configs/mot.yaml except we set pcl.G=20. Unfortunately, we lost the data and model weights in a disk failure and need some time to replicate them.

• [x] update training setting for Personpath.
• [ ] update training setting for KITTI.

Citation

@InProceedings{Lu_2024_CVPR,
    author    = {Lu, Zijia and Shuai, Bing and Chen, Yanbei and Xu, Zhenlin and Modolo, Davide},
    title     = {Self-Supervised Multi-Object Tracking with Path Consistency},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2024},
    pages     = {19016-19026}
}