Method

Cascade TWiX [on] [C-TWiX]
https://github.com/Guepardow/TWiX/

Submitted on 16 Nov. 2024 16:18 by
Mehdi Miah (Polytechnique Montréal)

Running time:0.01 s
Environment:8 cores @ >3.5 Ghz (Python)

Method Description:
No CMC, no visual appearances, no ReID, no BEV, no LIDAR, no depth estimation; only spatio-temporal coordinates (xmin, ymin, xmax, ymax, t).

Detections : Permatrack
Pipeline : Online Cascade matching (cf C-BIoU)

We focused on the association step by training a Transformer-based network to predict the similarity matrix between past tracks and current detections.
Parameters:
Maximal temporal gap at STA/LTA : 0.0s/0.8s
Past window size : 0.4s
Future window size : 1/fps sec
Matching association thresholds at STA/LTA : 0.4/-0.6

Tracking maximum age : 0.8s
Tracking Minimal score : 50%
Latex Bibtex:
@article{miah2024learningdata,
title = {Learning data association for multi-object tracking using only coordinates},
journal = {Pattern Recognition},
volume = {160},
pages = {111169},
year = {2025},
issn = {0031-3203},
doi = {https://doi.org/10.1016/j.patcog.2024.111169},
url = {https://www.sciencedirect.com/science/article/pii/S0031320324009208},
author = {Mehdi Miah and Guillaume-Alexandre Bilodeau and Nicolas Saunier},
keywords = {Tracking, Transformer, Data association, Motion, Multi-object tracking}
}

Detailed Results

From all 29 test sequences, our benchmark computes the commonly used tracking metrics CLEARMOT, MT/PT/ML, identity switches, and fragmentations [1,2]. The tables below show all of these metrics.


Benchmark MOTA MOTP MODA MODP
CAR 90.03 % 85.62 % 91.03 % 88.24 %
PEDESTRIAN 64.32 % 75.52 % 65.34 % 92.28 %

Benchmark recall precision F1 TP FP FN FAR #objects #trajectories
CAR 92.83 % 99.14 % 95.88 % 35907 313 2772 2.81 % 44173 1011
PEDESTRIAN 71.66 % 92.24 % 80.66 % 16732 1407 6617 12.65 % 21306 460

Benchmark MT PT ML IDS FRAG
CAR 82.15 % 14.92 % 2.92 % 344 620
PEDESTRIAN 42.61 % 39.86 % 17.53 % 236 896

This table as LaTeX

[1] K. Bernardin, R. Stiefelhagen: Evaluating Multiple Object Tracking Performance: The CLEAR MOT Metrics. JIVP 2008.
[2] Y. Li, C. Huang, R. Nevatia: Learning to associate: HybridBoosted multi-target tracker for crowded scene. CVPR 2009.


eXTReMe Tracker