Looking Beyond Two Frames: End-to-End Multi-Object Tracking Using Spatial and Temporal Transformers.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Multi-object tracking (MOT) faces challenges with occlusions and indefinite object numbers.
• Existing MOT methods often neglect crucial long-term temporal information, hindering performance.

Purpose of the Study:

• To present MO3TR, an end-to-end Transformer-based online multi-object tracking framework.
• To address limitations in handling occlusions and managing track initiation/termination in MOT.

Main Methods:

• MO3TR utilizes spatial and temporal Transformers to encode object interactions into long-term temporal embeddings.
• A spatial attention mechanism learns object-to-object and object-to-measurement representations.
• A temporal attention mechanism focuses on past information to resolve occlusions across frames.

Main Results:

• MO3TR achieves state-of-the-art or competitive results on popular multi-object tracking benchmarks.
• The framework demonstrates effective handling of occlusions, track initiation, and termination.
• Experiments validate the potential of the proposed Transformer-based approach.

Conclusions:

• MO3TR offers a novel, end-to-end solution for online multi-object tracking.
• The integration of spatial and temporal Transformers significantly improves occlusion handling and tracking accuracy.
• This framework advances the field of multi-object tracking by effectively utilizing long-term temporal dependencies.