Reproduce “Person Re-Identification with a Locally Aware Transformer” on OpenResearch

About this paper

Charu Sharma, Siddhant R. Kapil, David Chapman

This paper introduces the Locally Aware Transformer (LA-Transformer), a system that enhances person re-identification by effectively combining global and local features from Vision Transformers within a part-based architecture. The proposed method utilizes globally enhanced local tokens and a blockwise fine-tuning strategy to achieve state-of-the-art results on the Market-1501 and CUHK-03 benchmark datasets.

Key insights

• Combining globally enhanced local tokens from Vision Transformers significantly improves person re-identification performance compared to using only global or local features.
• A spatially aware, part-based processing strategy, inspired by traditional CNN methods, effectively leverages the spatial information within Vision Transformer's local tokens for fine-grained recognition.
• Blockwise fine-tuning is an effective strategy to adapt large, pre-trained Vision Transformers to smaller, domain-specific datasets like those for person re-identification, mitigating overfitting and improving convergence.

Summary by alphaXiv.