TransCenter: Transformers With Dense Representations for Multiple-Object Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Transformers demonstrate strong performance across various tasks.
• Their application to computer vision, including image classification and object detection, is growing.
• However, an accurate and efficient transformer-based multiple-object tracking (MOT) method is still needed.

Purpose of the Study:

• To design an accurate and efficient multiple-object tracking (MOT) method using transformer architecture.
• To address the limitations of direct transformer application in MOT, such as quadratic complexity and sparse queries.
• To propose TransCenter, a transformer-based MOT architecture optimized for accuracy and runtime.

Main Methods:

• Proposed TransCenter, a transformer-based MOT architecture employing dense representations.
• Introduced image-related dense detection queries for robust global target localization via heatmaps.
• Developed efficient sparse tracking queries from query learning networks (QLN) for temporal association.

Main Results:

• TransCenter achieved significant performance improvements, outperforming current state-of-the-art methods on standard MOT benchmarks.
• The method demonstrated remarkable accuracy and efficiency in both public and private tracking settings.
• Ablation studies and comparisons confirmed TransCenter's effectiveness against alternative approaches.

Conclusions:

• TransCenter offers an effective transformer-based solution for multiple-object tracking (MOT).
• The architecture's dense representation and novel query strategies enable accurate and efficient object tracking.
• The proposed method sets a new benchmark for MOT performance and efficiency.