Merlin2013 / TubeTK

Official implementation of paper: TubeTK: Adopting Tubes to Track Multi-Object in a One-Step Training Model (CVPR 2020 oral)

TubeTK

TubeTK is an one-step end-to-end multi-object tracking method, which is the first end-to-end open-source system that achieves 60+ MOTA on MOT-16 (64 MOTA) and MOT-17 (63 MOTA) datasets. Our paper "TubeTK: Adopting Tubes to Track Multi-Object in a One-Step Training Model" is accepted as an oral paper on CVPR-2020.

Contents

• TubeTK
• Contents
• Results
  • MOT-16
  • MOT-17
• Installation
• Quick Start
  • Demo
  • Evaluation
  • Train

Results

MOT-16

Results on MOT-16 dataset:

Video	MOTA	IDF1	MT	ML	FP	FN	IDS
MOT16-01	48.9	45.5	8	9	175	3052	40
MOT16-03	76.3	69.5	86	12	3741	20828	177
MOT16-06	51.2	55.7	87	39	1863	3542	231
MOT16-07	55.0	43.5	21	3	2225	4938	190
MOT16-08	46.9	37.3	18	3	1694	6952	234
MOT16-12	52.4	50.8	27	20	533	3366	51
MOT16-14	35.8	39.8	7	61	731	10948	194
TubeTK (Mean)	64.0	59.4	33.5	19.4	10962	53626	1117
RAN	63.0	63.8	39.9	22.1	13663	53248	482
Tracktor	54.5	52.5	19.0	36.9	3280	79149	682

MOT-17

Results on MOT-17 dataset:

Video	MOTA	IDF1	MT	ML	FP	FN	IDS
MOT17-01	47.9	44.9	6	10	167	3154	41
MOT17-03	76.4	69.6	81	12	3181	21287	186
MOT17-06	52.4	54.8	85	36	1609	3699	307
MOT17-07	55.4	43.3	21	2	1944	5371	222
MOT17-08	42.3	34.1	18	12	970	10889	319
MOT17-12	50.3	49.4	28	23	494	3749	63
MOT17-14	35.6	39.5	6	61	655	11012	241
TubeTK (Mean)	63.0	58.6	31.2	19.9	27060	177483	4137
SCNet	60.0	54.4	34.4	16.2	72230	145851	7611
Tracktor	53.5	52.3	19.5	36.3	12201	248047	2072

Installation

1. Get the code and build related modules:

     git clone ...(TO BE CONFIRM)
     cd TubeTK/install
     ./compile.sh
     # if something wrong, try:
     # sudo ldconfig <path/to/cuda>/lib64
     cd ..

2. Install pytorch 1.10 and other dependencies:

pip install -r requirements.txt

3. If the memory of your GPU < 16G, then you need NVIDIA APEX to conduct the mixed precision training.

   1. Install Apex:

   git clone https://github.com/NVIDIA/apex
   cd apex
   pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
   # if something wrong with the above pip install, try:
   # pip install -v --no-cache-dir ./

   2. We provide the --apex option to train with the APEX, see Quick Start for detail.

4. Run fetch_model.sh to download our pre-trained models. Or download the models manually and put them in ./models:

   1. 3DResNet50_original (Baidu pan | Google drive)

Quick Start

Demo

Run TubeTK for a video and visualization the results with:

python launch.py --nproc_per <num of GPU> --training_script demo.py --batch_size=3 --config configs/TubeTK_resnet_50_FPN_8frame_1stride.yaml --video_url <folder/to/the/videos> --output_dir ./vis_video

Evaluation on MOT-17 (16)

1. Download the data from MOT Challenge, and put or link it to ./data

2. To get the tracking result with:

   python launch.py --nproc_per <num of GPU> --training_script evaluate.py --batch_size 3 --config configs/TubeTK_resnet_50_FPN_8frame_1stride.yaml --trainOrTest test

3. To get the visualization with:

   python Visualization/Vis_Res.py --mode test

   The visualization videos are stored in ./vis_video .

Train on MOT-17 (16)

1. Download the data from MOT Challenge, and put or link it to ./data

2. Get the ground truth Btubes with:

   python ./pre_processing/get_tubes_MOT17.py

3. Train the model with:

   python launch.py --nproc_per <num of GPU> --training_script main.py --batch_size 1 --config ./configs/TubeTK_resnet_50_FPN_8frame_1stride.yaml

   If out of memory, try:

   python launch.py --nproc_per <num of GPU> --training_script main.py --batch_size 1 --config ./configs/TubeTK_resnet_50_FPN_8frame_1stride.yaml --apex

   If still out of memory, modify the configuration file: TubeTK_resnet_50_FPN_8frame_1stride.yaml:

   tube_limit: 500  # or 300
   

Citation

@inproceedings{pang2020tubetk,
   title={TubeTK: Adopting Tubes to Track Multi-Object in a One-Step Training Model},
   author={Pang, Bo and Li, Yizhuo and Zhang, Yifan and Li, Muchen and Lu, Cewu},
   booktitle={CVPR},
   year={2020}
}

License

TubeTK is freely available for free non-commercial use, and may be redistributed under these conditions.