SLATracker
This repository hosts our code for our paper Spatial-Attention Location-Aware Multi-Object Tracking. (arxiv link not available now)
The code will be released after the paper published.
Demo Bilibili
Requirements
- Python3.6
- Pytorch 1.6.0, torchvision 0.7.0
- detectron2
- python-opencv
- py-motmetrics
- cython-bbox
Installation
See INSTALL.md
Quick Start
Dataset Zoo
See DATASET_ZOO.md
Note that we transform all used datasets to COCO format for training convenience.
The pre-processing code will be uploaded later. Or you can download our transformed .json files from GoogleDrive
Training
python3 train_net.py --config-file configs/faster_rcnn_R_50_FPN_1x.yaml
Visualization
A trained model is available at GoogleDrive or BaiduDisk (kw:sucy)
python3 demo/vis_track.py --config-file configs/faster_rcnn_R_50_FPN_1x.yaml --opts MODEL.WEIGHTS output/model_final.pth
MOTChallenge Results
Public
| Benchmark | MOTA | IDF1 | HOTA | MOTP | MT | ML | FP | FN | IDSw |
|---|---|---|---|---|---|---|---|---|---|
| 2DMOT15 | 47.0 | 57.9 | 43.0 | 75.3 | 22.6 | 27.2 | 9044 | 22986 | 558 |
| MOT16 | 60.6 | 59.5 | 46.8 | 78.0 | 24.2 | 29.1 | 5783 | 65469 | 643 |
| MOT17 | 59.7 | 63.4 | 49.1 | 77.7 | 24.0 | 31.1 | 16644 | 209318 | 1647 |
Note that we utilize the pre-processing method of Tracktor, not CenterTrack.
Private
| Benchmark | MOTA | IDF1 | HOTA | MOTP | MT | ML | FP | FN | IDSw |
|---|---|---|---|---|---|---|---|---|---|
| 2DMOT15 | 57.9 | 62.2 | 47.1 | 75.8 | 39.1 | 14.7 | 6973 | 18313 | 577 |
| MOT16 | 72.0 | 69.6 | 54.7 | 77.9 | 37.3 | 20.9 | 7242 | 43147 | 740 |
| MOT17 | 71.8 | 69.0 | 54.4 | 77.8 | 38.0 | 20.5 | 19077 | 137700 | 2493 |
Acknowledgement
A large part of the code is borrowed from Zhongdao/Towards-Realtime-MOT and DeanChan/HOIM-PyTorch. Thanks for their wonderful works.