Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

BoostTrack++: using tracklet information to detect more objects in multiple object tracking

Created by
  • Haebom

Author

Vuka\v{s}in Stanojevi c, Branimir Todorovi c

Outline

This paper focuses on the problem of selecting high-confidence detection bounding boxes in Multi-Object Tracking (MOT). While the existing BoostTrack approach attempts to overcome the shortcomings of multi-level linkage methods by improving detection confidence, this paper addresses the limitations of BoostTrack's confidence enhancement techniques and proposes a novel approach to improve performance. The proposed method combines shape, Mahalanobis distance, and a novel soft BIoU similarity measure to construct a richer similarity measure and improve the selection of true positive detections. Furthermore, we introduce a soft detection confidence enhancement technique that computes a new confidence score based on the similarity measure and previous confidence scores, and a variable similarity threshold to account for irregularly updated low similarity measures between detections and tracklets. The proposed additions are independent and applicable to any MOT algorithm. Consequently, combining the proposed method with the BoostTrack+ baseline achieves results approaching the state-of-the-art on the MOT17 dataset and new state-of-the-art results on the HOTA and IDF1 scores on the MOT20 dataset. Source code is provided.

Takeaways, Limitations

•
Takeaways:
◦
An effective solution to the problem of selecting reliable detection bounding boxes in multi-object tracking (MOT).
◦
A novel similarity measure combining shape, Mahalanobis distance, and soft BIoU similarity is proposed.
◦
Performance improvement through soft detection reliability enhancement techniques and variable similarity threshold introduction.
◦
Achieving state-of-the-art performance on MOT17 and MOT20 datasets.
◦
The proposed method is modularized and can be easily applied to other MOT algorithms.
•
Limitations:
◦
The performance improvement of the proposed method may be limited to specific datasets.
◦
Further validation of generalization performance in various environments and conditions is needed.
◦
A more detailed comparative analysis with other state-of-the-art MOT algorithms may be needed.
👍
Made with Slashpage