#sportstrack
This repository is an official implementation of the SportsTrack: An Innovative Method for Tracking Athletes in Sports Scenes
1st place of 2022 ECCV Sports MOT challenge.
| Method | HOTA | AssA | DetA | MOTA | weight |
|---|---|---|---|---|---|
| SportsTrack | 76.264 | 73.538 | 79.180 | 89.316 | baidu/onedrive |
-
Download SportsMOT dataset from here
-
set data root path of sports MOT dataset
export SPORTSTRACK_ROOT=...
ls ${SPORTSTRACK_ROOT}/data/sportsmot_publish/dataset/test
output
v_1UDUODIBSsc_c001 v_2ChiYdg5bxI_c058 v_2Dw9QNH5KtU_c014 v_7FTsO8S3h88_c007 v_9p0i81kAEwE_c010 v_BdD9xu0E2H4_c011 v_czYZnO9QxYQ_c020
...
- build gmc
build gmc need opencv4
cd gmc
mkdir build
cd build
cmake ..
make
./gmc ${SPORTSTRACK_ROOT}/data/sportsmot_publish/dataset/test/* #this will generate a gmc.txt file in each input directorys
./gmc ${SPORTSTRACK_ROOT}/data/sportsmot_publish/dataset/val/*
- run algorithm
python main.py --split val --gpus 0 #track on val dataset
or
python main.py --split test --gpus 0 #track on test dataset
or use multi process
python multi_process_main.py --split test --gpus 0 #use multiprocess to track on test dataset
- post process
python postprocess_tools/merge_tracks.py --image_dir ... --input_txt_dir ... --output_txt_dir ...
python postprocess_tools/interpolation.py --input_txt_dir ... --output_txt_dir ...
Example:
python postprocess_tools/merge_tracks.py --image_dir ${SPORTSTRACK_ROOT}/data/sportsmot_publish/dataset/test --input_txt_dir ${SPORTSTRACK_ROOT}/tmp/sportsmot-test/PDSMV2SportsTrackerT1/data/ --output_txt_dir ${SPORTSTRACK_ROOT}/tmp/sportsmot-test/PDSMV2SportsTrackerT1/data_merge
python postprocess_tools/interpolation.py --input_txt_dir ${SPORTSTRACK_ROOT}/tmp/sportsmot-test/PDSMV2SportsTrackerT1/data_merge/ --output_txt_dir ${SPORTSTRACK_ROOT}/tmp/sportsmot-test/PDSMV2SportsTrackerT1/data_inter
- pytorch
- wml
install wml
git clone [email protected]:vghost2008/wml.git
cd wml
export PYTHONPATH=${PYTHONPATH}:`pwd`
- Jie Wang
- Xiaodong Yang
- YuZhou Peng
- Ting Wang
- Yanming Zhang
We acknowledge the excellent implementation from YOLOX , FastReID and HRNet.