Evaluation Code for - MTMC Tracking 2025 Dataset

Evaluation code for the Multi-Target Multi-Camera (MTMC) 2025 dataset.

The evaluation utilizes Higher Order Tracking Accuracy (HOTA) score as an evaluation metric for multi-object tracking that addresses the limitations of previous metrics like MOTA and IDF1. It integrates three key aspects of MOT: accurate detection, association, and localization into a unified metric. This comprehensive approach balances the importance of detecting each object (detection), correctly identifying objects across different frames (association), and accurately localizing objects in each frame (localization). Furthermore, HOTA can be decomposed into simpler components, allowing for detailed analysis of different aspects of tracking behavior.

HOTA scores calculated using 3D IOU of the bounding box dimension in a multi-camera setting.

Environment setup:

+ conda env create -f environment.yml
+ conda activate mtmc_eval_2025

Usage:

# The prediction file is a copy of the ground truth file used for testing

python3 main.py --ground_truth_file ./sample_data/ground_truth_test_full.txt --input_file sample_data/pred.txt --scene_id_2_scene_name_file sample_data/scene_id_2_scene_name_full.json --num_cores 16

Sample Result
--------------------------------------------------------------
25/07/11 21:01:47 - Final HOTA: 100.0
25/07/11 21:01:47 - Final DetA: 100.0
25/07/11 21:01:47 - Final AssA: 100.0
25/07/11 21:01:47 - Final LocA: 99.99993990354818
25/07/11 21:01:47 - Total time taken: 522.6667423248291 seconds

Note: A sample ground truth file is provided in sample_data/ground_truth_test_full.txt for demonstration purposes. Please replace this with your own ground truth file corresponding to the dataset split you are evaluating against.

Acknowledgements

This project utilizes a portion of code from TrackEval, an open-source project by Jonathon Luiten for evaluating multi-camera tracking results. TrackEval is licensed under the MIT License, which you can find in full here.