Faster Person Re-Identification: One-Shot-Filter and Coarse-to-Fine Search.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Fast person re-identification (ReID) is crucial for efficient image search.
• Current hashing-based ReID methods achieve speed but require long codes for accuracy, creating a speed-accuracy trade-off.
• Existing methods face scalability challenges with large datasets due to O(n) time complexity.

Purpose of the Study:

• To develop a novel hashing code search strategy for fast and accurate person ReID.
• To address the speed-accuracy trade-off in hashing-based ReID.
• To improve the efficiency and scalability of ReID methods for large-scale galleries.

Main Methods:

• Introduced a Coarse-to-Fine (CtF) hashing code search strategy using complementary short and long codes.
• Designed an All-in-One (AiO) module for simultaneous learning of multi-length codes and a Distance Threshold Optimization (DTO) algorithm for threshold searching.
• Proposed a gallery-size-free One-Shot-Filter (OSF) strategy with a Latent-Attribute-Learning (LAL) module and Single-Direction-Metric (SDM) Loss for O(1) complexity filtering.

Main Results:

• The CtF+OSF method achieved 2% higher accuracy and was 5x faster than contemporary hashing ReID methods.
• CtF demonstrated 50x speedup with comparable accuracy against non-hashing ReID methods.
• OSF further accelerated CtF by 2x, resulting in up to 10x total speedup with minimal accuracy loss.

Conclusions:

• The proposed CtF strategy effectively balances speed and accuracy by leveraging multi-length codes.
• The OSF strategy significantly enhances scalability and search speed for large-scale ReID tasks.
• The combined CtF+OSF approach represents a significant advancement in efficient and accurate person re-identification.