Relation-Based Deep Attention Network with Hybrid Memory for One-Shot Person Re-Identification

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Person re-identification (Re-ID) faces challenges due to limited labeled data.
• Existing methods struggle to maintain performance with scarce labeled samples.

Purpose of the Study:

• To develop an effective one-shot person Re-ID method using limited labeled data.
• To improve the performance of person Re-ID in low-data scenarios.

Main Methods:

• Proposed a relation-based attention network with hybrid memory for person Re-ID.
• Developed a specialized network architecture to reduce environmental noise interference.
• Introduced a hybrid memory framework for unified training of one-shot and unlabeled data.
• Implemented a one-shot feature update mode to mitigate overfitting.

Main Results:

• Achieved significant performance improvements on Market-1501, DukeMTMC-reID, and MSMT17 datasets.
• Demonstrated considerable gains of 6.7%, 4.6%, and 11.5% respectively.
• Established a new state-of-the-art for one-shot person Re-identification.

Conclusions:

• The proposed relation-based attention network with hybrid memory is highly effective for one-shot person Re-ID.
• The method successfully overcomes the limitations of insufficient labeled data.
• This approach sets a new benchmark in the field of person Re-identification.