Multiple Object Tracking from appearance by hierarchically clustering tracklets

Presented as a spotlight in BMVC22

Multiple Object Tracking from appearance by hierarchically clustering tracklets

Tracking performance

Results on test set of different datasets

Dataset	HOTA	DetA	AssA	MOTA	IDF1
MOT17	62.6	62.2	63.4	76.7	77.7
MOT20	57.3	56.7	58.1	68.0	69.7
DanceTrack	48.7	79.8	29.9	89.9	46.5

Installation

Code tested in python3.8

Download github repository and position in the ROOT folder

git clone https://github.com/NII-Satoh-Lab/MOT_FCG.git

cd $YOUR_PATH/MOT_FCG

Install dependencies

pip install -r requirements.txt

Data preparation

Download MOT17, MOT20, and DanceTrack.
Download detections and feature data (around 20GB)

Note: This scripts downloads and prepares detections and features for MOT17, MOT20, and dancetrack. If you want to download only a specific dataset modify the below script.

bash scripts/prepare_data.sh

Run example

Run the code example for MOT17 using YOLOX and SBS features

python src/fcg_tracking.py --gt $YOUR_PATH_TO_DATASETS/MOT17Det --dataset mot17 --mode train --detector YOLOX --reid SBS --temp --iou --spatial --motion --w_tracklet 6 --w_fuse 3 --max_prop 40

Evaluation

Download evaluation repository from TrackEval

Download the git repository and follow the installation instructions and the ground truth data positioning.

git clone https://github.com/JonathonLuiten/TrackEval

python $TRACKEVAL_PATH/scripts/run_mot_challenge.py --TRACKERS_FOLDER $YOUR_PATH/MOT_FCG/results/mot17 --BENCHMARK MOT17 --TRACKERS_TO_EVAL fcg-YOLOX-SBS --GT_FOLDER $TRACKEVAL_PATH/data/MOT17 --SPLIT_TO_EVAL train --METRICS HOTA CLEAR Identity --PLOT_CURVES False --PRINT_CONFIG False --TIME_PROGRESS False

Citation

@inproceedings{Girbau_2022_BMVC,
author    = {Andreu Girbau and Ferran Marques and Shin'ichi Satoh},
title     = {Multiple Object Tracking from appearance by hierarchically clustering tracklets},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year      = {2022},
url       = {https://bmvc2022.mpi-inf.mpg.de/0362.pdf}
}

TODO

Code with our extracted detections + features
Bibcitation
Script to generate visualizations (e.g. videos)
Scripts to generate low fps MOT17 videos