Pseudo-Pair Based Self-Similarity Learning for Unsupervised Person Re-Identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (re-ID) is crucial for video surveillance.
• Existing re-ID methods often require extensive labeled data for supervised training.
• Unsupervised approaches are needed to overcome data annotation limitations.

Purpose of the Study:

• To develop an unsupervised person re-ID method using pseudo-pair based self-similarity learning.
• To eliminate the need for human annotations in training re-ID models.
• To improve the generalization of similarity functions in re-ID tasks.

Main Methods:

• Proposed a pseudo-pair based self-similarity learning approach for unsupervised person re-ID.
• Constructed patch surrogate classes for initial supervision and used gradient-guided similarity separation for pseudo-labeling.
• Introduced intra-similarity learning (channel attention) for local features and inter-similarity learning (deformable convolution, non-local block) for patch correspondence.

Main Results:

• The proposed method achieves superior performance on several person re-ID benchmark datasets.
• Demonstrated the effectiveness of the unsupervised approach without human annotations.
• Outperformed existing state-of-the-art methods in unsupervised person re-ID.

Conclusions:

• The novel pseudo-pair based self-similarity learning approach is effective for unsupervised person re-ID.
• The method successfully learns discriminative local features and cross-image patch correspondence.
• This work offers a promising direction for efficient and accurate person re-identification in real-world surveillance.