Learning Occlusion-Dynamic Invariant Representations for Multi-Object Tracking

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Multi-object tracking (MOT) performance degrades due to unstable appearance features under visual corruptions.
• Occlusion and motion blur introduce noise, weakening temporal representations and leading to identity switches.

Purpose of the Study:

• To develop a method for learning more stable appearance representations resilient to feature corruption in online tracking.
• To improve the robustness of multi-object tracking, particularly in challenging visual conditions.

Main Methods:

• Proposes the Causal Interaction Module (CIM), a causal architecture with a filter-then-reconstruct design.
• A temporal filtering stage creates a stable anchor from historical feature trajectories.
• A contextual enhancement stage refines frame-level features using the anchor for improved association.

Main Results:

• CIM integration into standard trackers enhances association robustness.
• Consistent performance gains observed across multiple MOT benchmarks.
• Significant improvements noted on association-related metrics, especially under corruption stress tests.

Conclusions:

• The Causal Interaction Module (CIM) effectively addresses appearance feature instability in multi-object tracking.
• CIM enhances tracking robustness without altering the core tracking formulation.
• The proposed method shows promise for improving MOT in real-world scenarios with visual challenges.