Inference-Domain Network Evolution: A New Perspective for One-Shot Multi-Object Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Supervised one-shot multi-object tracking (MOT) relies heavily on extensive manual annotations.
• Real-world applications face challenges in acquiring sufficient labeled data for MOT models.
• Domain adaptation for MOT is difficult due to discrepancies in object appearance, identity, quantity, and scale across domains.

Purpose of the Study:

• To enhance the generalization ability of one-shot MOT models for unlabeled domains.
• To develop a method for adapting pre-trained MOT models without requiring new manual annotations.
• To address the challenges of domain shift in multi-object tracking tasks.

Main Methods:

• Proposed a spatial topology-based one-shot network (STONet) incorporating a self-supervision mechanism for learning spatial contexts.
• Introduced a temporal identity aggregation (TIA) module to mitigate noisy labels during network evolution.
• Implemented an inference-domain network evolution strategy for progressive pseudo-label collection and parameter updates.

Main Results:

• The proposed STONet with TIA demonstrated effective domain adaptation for one-shot MOT.
• Self-supervision enabled feature extractors to learn spatial contexts without annotated data.
• The TIA module improved the reliability of pseudo-labels, enhancing model performance in the target domain.

Conclusions:

• The novel inference-domain network evolution approach significantly improves the generalization of one-shot MOT models.
• The combination of STONet and TIA offers a practical solution for MOT in unlabeled or cross-domain scenarios.
• The method effectively bridges the gap between labeled source domains and unlabeled target domains in MOT.