[update 20231122]
fix reid FP16 bug in my RP to official ByteTrack. Now ReID module brings more befifits.
Also support HOTA metric from TrackEval.
| Tracker | before fix ReID FP 16 Bug AP50:95 AP50 AP75 | HOTA MOTA IDF1 |
after fix ReID FP 16 Bug AP50:95 AP50 AP75 | HOTA MOTA IDF1 |
|---|---|---|
| ByteTrack | 0.567 0.868 0.647 | 62.038 73.297 74.935 | 0.567 0.868 0.647 | 62.038 73.297 74.935 |
| FairMOT | 0.567 0.868 0.647 | 61.921 73.292 74.754 | 0.567 0.868 0.647 | 61.921 73.292 74.754 |
| FairMOT+BYTE | 0.567 0.868 0.647 | 61.602 73.359 73.648 | 0.567 0.868 0.647 | 61.602 73.359 73.648 |
[update 20220514]:
Onedrive link of trained model. Trained model's mAP should be 0.556 with MOTA 72.6 using ByteTrack and MOTA 70.9 using FairMOT tracking strategy.
[update 20220511]:
-
Please open issue with English, not Chinese, so the discussion can benifit the community.
-
To switch from different trackers to another, please replace the byte_track.py with other files.
-
Training proceduer is the same as original ByteTrack. If you want to train the model on larger datasets with ids, please follow JDE/FairMOT.
[update 20220428]:
I found a ReID related bug of original ByteTrack. I made a PR to ByteTrack and it is merged to master branch of ByteTrack. ifzhang/ByteTrack#184
So the ReID part of current code will not be train correctly when track_id becomes larger. I will update the code when I got time.
Or you can make a PR to help me out!
[update 20220414]:
- Fix loss computation bug in yolo_head.py.
- Fix feature update in FairMOT tracker when recovering a tracklet.
- Fix training set in yolox_s_mot_half using train_half.json instead of train.json.
- Trained model can be download here jc69. Trained model's mAP should be 0.556 with MOTA 72.6 using ByteTrack and MOTA 70.9 using FairMOT tracking strategy.
- Note that reid embeddings only trained on MOT17 half is not reliable due to limited ID annotations.
ByteTrack is the SOTA tracker in MOT benchmarks with strong detector YOLOX and a simple association strategy only based on motion information.
Motion information (IoU distance) is efficient and effective in short-term tracking, but can not be used for recovering targets after long-time disappear or conditions with moving camera.
So it is important to enhance ByteTrack with a ReID module for long-term tracking, improving the performance under other challenging conditions, such as moving camera.
Some code is borrowed from FairMOT
For now, the results are trained on half of MOT17 and tested on the other half of MOT17. And the performance is lower than the original performance.
Any issue and suggestions are welcome!
- Enhanced ByteTrack with a ReID module (head) following the paradigm of FairMOT.
- Add a classifier for supervised training of ReID head.
- Using uncertainty loss in FairMOT for the balance of detection and ReID tasks.
- Tracking strategy is borrowed from FairMOT
- support more datasets
- single class –> multiple class
- other loss functions for better ReID performance
- other strategies for multiple tasks balance
- … …
The following contents is original README in ByteTrack.
ByteTrack: Multi-Object Tracking by Associating Every Detection Box
Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Zehuan Yuan, Ping Luo, Wenyu Liu, Xinggang Wang
| Google Colab demo | Huggingface Demo | Original Paper: ByteTrack |
|---|---|---|
 |
 |
arXiv 2110.06864 |
- Integrated to Huggingface Spaces with Gradio.
Multi-object tracking (MOT) aims at estimating bounding boxes and identities of objects in videos. Most methods obtain identities by associating detection boxes whose scores are higher than a threshold. The objects with low detection scores, e.g. occluded objects, are simply thrown away, which brings non-negligible true object missing and fragmented trajectories. To solve this problem, we present a simple, effective and generic association method, tracking by associating every detection box instead of only the high score ones. For the low score detection boxes, we utilize their similarities with tracklets to recover true objects and filter out the background detections. When applied to 9 different state-of-the-art trackers, our method achieves consistent improvement on IDF1 scores ranging from 1 to 10 points. To put forwards the state-of-the-art performance of MOT, we design a simple and strong tracker, named ByteTrack. For the first time, we achieve 80.3 MOTA, 77.3 IDF1 and 63.1 HOTA on the test set of MOT17 with 30 FPS running speed on a single V100 GPU.
| Dataset | MOTA | IDF1 | HOTA | MT | ML | FP | FN | IDs | FPS |
|---|---|---|---|---|---|---|---|---|---|
| MOT17 | 80.3 | 77.3 | 63.1 | 53.2% | 14.5% | 25491 | 83721 | 2196 | 29.6 |
| MOT20 | 77.8 | 75.2 | 61.3 | 69.2% | 9.5% | 26249 | 87594 | 1223 | 13.7 |
Step1. Install ByteTrack.
git clone https://github.com/ifzhang/ByteTrack.git
cd ByteTrack
pip3 install -r requirements.txt
python3 setup.py developStep2. Install pycocotools.
pip3 install cython; pip3 install 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'Step3. Others
pip3 install cython_bboxdocker build -t bytetrack:latest .
# Startup sample
mkdir -p pretrained && \
mkdir -p YOLOX_outputs && \
xhost +local: && \
docker run --gpus all -it --rm \
-v $PWD/pretrained:/workspace/ByteTrack/pretrained \
-v $PWD/datasets:/workspace/ByteTrack/datasets \
-v $PWD/YOLOX_outputs:/workspace/ByteTrack/YOLOX_outputs \
-v /tmp/.X11-unix/:/tmp/.X11-unix:rw \
--device /dev/video0:/dev/video0:mwr \
--net=host \
-e XDG_RUNTIME_DIR=$XDG_RUNTIME_DIR \
-e DISPLAY=$DISPLAY \
--privileged \
bytetrack:latestDownload MOT17, MOT20, CrowdHuman, Cityperson, ETHZ and put them under <ByteTrack_HOME>/datasets in the following structure:
datasets
|——————mot
| └——————train
| └——————test
└——————crowdhuman
| └——————Crowdhuman_train
| └——————Crowdhuman_val
| └——————annotation_train.odgt
| └——————annotation_val.odgt
└——————MOT20
| └——————train
| └——————test
└——————Cityscapes
| └——————images
| └——————labels_with_ids
└——————ETHZ
└——————eth01
└——————...
└——————eth07
Then, you need to turn the datasets to COCO format and mix different training data:
cd <ByteTrack_HOME>
python3 tools/convert_mot17_to_coco.py
python3 tools/convert_mot20_to_coco.py
python3 tools/convert_crowdhuman_to_coco.py
python3 tools/convert_cityperson_to_coco.py
python3 tools/convert_ethz_to_coco.pyBefore mixing different datasets, you need to follow the operations in mix_xxx.py to create a data folder and link. Finally, you can mix the training data:
cd <ByteTrack_HOME>
python3 tools/mix_data_ablation.py
python3 tools/mix_data_test_mot17.py
python3 tools/mix_data_test_mot20.pyTrain on CrowdHuman and MOT17 half train, evaluate on MOT17 half val
| Model | MOTA | IDF1 | IDs | FPS |
|---|---|---|---|---|
| ByteTrack_ablation [google], [baidu(code:eeo8)] | 76.6 | 79.3 | 159 | 29.6 |
Train on CrowdHuman, MOT17, Cityperson and ETHZ, evaluate on MOT17 train.
- Standard models
| Model | MOTA | IDF1 | IDs | FPS |
|---|---|---|---|---|
| bytetrack_x_mot17 [google], [baidu(code:ic0i)] | 90.0 | 83.3 | 422 | 29.6 |
| bytetrack_l_mot17 [google], [baidu(code:1cml)] | 88.7 | 80.7 | 460 | 43.7 |
| bytetrack_m_mot17 [google], [baidu(code:u3m4)] | 87.0 | 80.1 | 477 | 54.1 |
| bytetrack_s_mot17 [google], [baidu(code:qflm)] | 79.2 | 74.3 | 533 | 64.5 |
- Light models
| Model | MOTA | IDF1 | IDs | Params(M) | FLOPs(G) |
|---|---|---|---|---|---|
| bytetrack_nano_mot17 [google], [baidu(code:1ub8)] | 69.0 | 66.3 | 531 | 0.90 | 3.99 |
| bytetrack_tiny_mot17 [google], [baidu(code:cr8i)] | 77.1 | 71.5 | 519 | 5.03 | 24.45 |
Train on CrowdHuman and MOT20, evaluate on MOT20 train.
| Model | MOTA | IDF1 | IDs | FPS |
|---|---|---|---|---|
| bytetrack_x_mot20 [google], [baidu(code:3apd)] | 93.4 | 89.3 | 1057 | 17.5 |
The COCO pretrained YOLOX model can be downloaded from their model zoo. After downloading the pretrained models, you can put them under <ByteTrack_HOME>/pretrained.
- Train ablation model (MOT17 half train and CrowdHuman)
cd <ByteTrack_HOME>
python3 tools/train.py -f exps/example/mot/yolox_x_ablation.py -d 8 -b 48 --fp16 -o -c pretrained/yolox_x.pth- Train MOT17 test model (MOT17 train, CrowdHuman, Cityperson and ETHZ)
cd <ByteTrack_HOME>
python3 tools/train.py -f exps/example/mot/yolox_x_mix_det.py -d 8 -b 48 --fp16 -o -c pretrained/yolox_x.pth- Train MOT20 test model (MOT20 train, CrowdHuman)
For MOT20, you need to clip the bounding boxes inside the image.
Add clip operation in line 134-135 in data_augment.py, line 122-125 in mosaicdetection.py, line 217-225 in mosaicdetection.py, line 115-118 in boxes.py.
cd <ByteTrack_HOME>
python3 tools/train.py -f exps/example/mot/yolox_x_mix_mot20_ch.py -d 8 -b 48 --fp16 -o -c pretrained/yolox_x.pth- Train custom dataset
First, you need to prepare your dataset in COCO format. You can refer to MOT-to-COCO or CrowdHuman-to-COCO. Then, you need to create a Exp file for your dataset. You can refer to the CrowdHuman training Exp file. Don't forget to modify get_data_loader() and get_eval_loader in your Exp file. Finally, you can train bytetrack on your dataset by running:
cd <ByteTrack_HOME>
python3 tools/train.py -f exps/example/mot/your_exp_file.py -d 8 -b 48 --fp16 -o -c pretrained/yolox_x.pth- Evaluation on MOT17 half val
Run ByteTrack:
cd <ByteTrack_HOME>
python3 tools/track.py -f exps/example/mot/yolox_x_ablation.py -c pretrained/bytetrack_ablation.pth.tar -b 1 -d 1 --fp16 --fuseYou can get 76.6 MOTA using our pretrained model.
Run other trackers:
python3 tools/track_sort.py -f exps/example/mot/yolox_x_ablation.py -c pretrained/bytetrack_ablation.pth.tar -b 1 -d 1 --fp16 --fuse
python3 tools/track_deepsort.py -f exps/example/mot/yolox_x_ablation.py -c pretrained/bytetrack_ablation.pth.tar -b 1 -d 1 --fp16 --fuse
python3 tools/track_motdt.py -f exps/example/mot/yolox_x_ablation.py -c pretrained/bytetrack_ablation.pth.tar -b 1 -d 1 --fp16 --fuse- Test on MOT17
Run ByteTrack:
cd <ByteTrack_HOME>
python3 tools/track.py -f exps/example/mot/yolox_x_mix_det.py -c pretrained/bytetrack_x_mot17.pth.tar -b 1 -d 1 --fp16 --fuse
python3 tools/interpolation.pySubmit the txt files to MOTChallenge website and you can get 79+ MOTA (For 80+ MOTA, you need to carefully tune the test image size and high score detection threshold of each sequence).
- Test on MOT20
We use the input size 1600 x 896 for MOT20-04, MOT20-07 and 1920 x 736 for MOT20-06, MOT20-08. You can edit it in yolox_x_mix_mot20_ch.py
Run ByteTrack:
cd <ByteTrack_HOME>
python3 tools/track.py -f exps/example/mot/yolox_x_mix_mot20_ch.py -c pretrained/bytetrack_x_mot20.pth.tar -b 1 -d 1 --fp16 --fuse --match_thresh 0.7 --mot20
python3 tools/interpolation.pySubmit the txt files to MOTChallenge website and you can get 77+ MOTA (For higher MOTA, you need to carefully tune the test image size and high score detection threshold of each sequence).
See tutorials.
Suppose you have already got the detection results 'dets' (x1, y1, x2, y2, score) from other detectors, you can simply pass the detection results to BYTETracker (you need to first modify some post-processing code according to the format of your detection results in byte_tracker.py):
from yolox.tracker.byte_tracker import BYTETracker
tracker = BYTETracker(args)
for image in images:
dets = detector(image)
online_targets = tracker.update(dets, info_imgs, img_size)
You can get the tracking results in each frame from 'online_targets'. You can refer to mot_evaluators.py to pass the detection results to BYTETracker.
cd <ByteTrack_HOME>
python3 tools/demo_track.py video -f exps/example/mot/yolox_x_mix_det.py -c pretrained/bytetrack_x_mot17.pth.tar --fp16 --fuse --save_result@article{zhang2021bytetrack,
title={ByteTrack: Multi-Object Tracking by Associating Every Detection Box},
author={Zhang, Yifu and Sun, Peize and Jiang, Yi and Yu, Dongdong and Yuan, Zehuan and Luo, Ping and Liu, Wenyu and Wang, Xinggang},
journal={arXiv preprint arXiv:2110.06864},
year={2021}
}
@article{zhang2021fairmot,
title={Fairmot: On the fairness of detection and re-identification in multiple object tracking},
author={Zhang, Yifu and Wang, Chunyu and Wang, Xinggang and Zeng, Wenjun and Liu, Wenyu},
journal={International Journal of Computer Vision},
volume={129},
pages={3069--3087},
year={2021},
publisher={Springer}
}
A large part of the code is borrowed from YOLOX, FairMOT, TransTrack and JDE-Cpp. Many thanks for their wonderful works.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent