Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

34 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

🎩 History-Aware Transformation of ReID Features for Multiple Object Tracking

This is the official PyTorch implementation of our paper:

History-Aware Transformation of ReID Features for Multiple Object Tracking
🎓 Ruopeng Gao, Yuyao Wang, Chunxu Liu, Limin Wang
📧 Primary contact: ruopenggao@gmail.com

🔍 Overview

TL; DR. We propose a plug-and-play History-Aware Transformation algorithm for appearance features (i.e., ReID features). Guided by historical information, we seek a more discriminative subspace for target features in video sequences, enabling better differentiation between different trajectories. Our method significantly enhances the reliability of appearance features, improving the performance of ReID-based MOT trackers.

overview

🔥 News

  • 2026.07.22: The latest version of our paper is now available on arXiv. It includes a more systematic analysis of the relationship between similarity measures and tracking performance, revealing long-standing issues in the ReID space and explaining how our method improves its representations. 🔥
  • 2026.07.18: The updated version of the paper will be posted on arXiv soon, with a more comprehensive analysis. 🔜
  • 2026.06.18: Our paper is accepted by ECCV 2026 ~ 🎉 🎉

📂 Sub-Projects

Theoretically, our proposed method can be applied to any multiple object tracker equipped with a ReID or appearance branch. Since different methods are often built on varying codebases, our repository is organized into separate sub-projects to ensure compatibility with different frameworks. Each sub-project should be treated and run independently.

Currently, we provide the following sub-project implementations:

  • HAT-MASA: MASA is a tracking model based purely on appearance feature matching, which has achieved remarkable performance in zero-shot tracking.
  • HAT-SORT: Since nearly all heuristic Multiple Object Tracking (MOT) algorithms adhere to the SORT-like design paradigm, we have selected a set of the most popular codebases as our foundation for implementation and modification.

✏️ Citation

If you think this project is helpful, please feel free to leave a ⭐ and cite our paper:

@article{{HATReID-MOT},
  title={History-aware transformation of reid features for multiple object tracking},
  author={Gao, Ruopeng and Wang, Yuyao and Liu, Chunxu and Wang, Limin},
  journal={arXiv preprint arXiv:2503.12562},
  year={2025}
}

Releases

Packages

Used by

Contributors

Languages