AAformer: Auto-Aligned Transformer for Person Re-Identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Accurate person re-identification (re-ID) relies on fine-grained, part-level features from images.
• Existing Convolutional Neural Network (CNN)-based methods often coarsely locate human parts or fail to identify non-human parts like accessories.
• Pretrained human parsing models have limitations in adaptability and identifying non-human elements crucial for re-ID.

Purpose of the Study:

• To propose a novel transformer-based architecture for automatic part localization in person re-ID.
• To introduce an alignment scheme that enables transformers to identify both human and non-human parts at a patch level.
• To enhance the feature extraction process for improved person re-identification performance.

Main Methods:

• Introduced the Auto-Aligned Transformer (AAformer) incorporating an alignment scheme within the transformer architecture.
• Developed 'Part tokens' ([PART]s) as learnable vectors to extract part-level features through localized self-attention.
• Implemented an auto-alignment mechanism using a fast optimal transport (OT) algorithm to cluster image patches around [PART] prototypes.

Main Results:

• AAformer successfully and automatically locates both human and non-human parts at the patch level.
• The proposed 'Part tokens' effectively capture discriminative part features for retrieval.
• Experimental results demonstrate the superiority of AAformer over existing state-of-the-art person re-identification methods.

Conclusions:

• The integration of part alignment into transformer self-attention is effective for person re-ID.
• AAformer offers a robust solution for extracting fine-grained features by automatically localizing relevant image parts.
• The proposed method advances the state-of-the-art in person re-identification by addressing limitations of previous approaches.