Learning Feature Recovery Transformer for Occluded Person Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (Re-ID) is significantly challenged by occlusions, leading to feature matching noise and information loss.
• Existing methods struggle to effectively handle partial or occluded pedestrian data.

Purpose of the Study:

• To propose a novel approach, the Feature Recovery Transformer (FRT), to address simultaneous challenges in occluded person Re-ID.
• To improve the accuracy and robustness of person Re-ID systems in the presence of occlusions.

Main Methods:

• Developed a visibility graph matching approach to focus on visible regions and reduce noise interference.
• Introduced a feature recovery transformer that leverages k-nearest neighbors' features to reconstruct lost pedestrian information.

Main Results:

• FRT demonstrated significant improvements across various person Re-ID datasets, including occluded, partial, and holistic ones.
• Achieved state-of-the-art performance on the Occluded-Duke dataset, with at least 6.2% higher Rank-1 accuracy and 7.2% higher mAP.

Conclusions:

• The Feature Recovery Transformer (FRT) effectively tackles noise interference and information loss in occluded person Re-ID.
• FRT offers a robust solution for person Re-ID tasks, particularly in real-world scenarios with frequent occlusions.