Disentangled Feature Learning Network and a Comprehensive Benchmark for Vehicle Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning
• Intelligent Transportation Systems

Background:

• Vehicle Re-Identification (ReID) is critical for public security and intelligent transportation.
• Existing ReID datasets suffer from constrained conditions, limited variations, and small scales.
• A large-scale, diverse dataset is essential for robust ReID model development and evaluation.

Purpose of the Study:

• Introduce VERI-Wild 2.0, a new large-scale benchmark dataset for vehicle ReID.
• Address the limitations of current datasets in terms of scale, diversity, and capture conditions.
• Propose a novel method to improve vehicle ReID performance by considering orientation information.

Main Methods:

• Collected 825,042 images from a city-scale surveillance system (274 cameras, >200 km²).
• Ensured rich appearance diversity through varied time spans, viewpoints, illumination, backgrounds, and weather.
• Developed a disentangled feature learning network (DFNet) to learn orientation-specific and orientation-invariant features.

Main Results:

• VERI-Wild 2.0 provides a challenging benchmark with a large training and non-overlapping test set (~400K images).
• The proposed DFNet effectively leverages orientation information for improved feature matching.
• Experimental results demonstrate the significant effectiveness of the DFNet for vehicle ReID.

Conclusions:

• VERI-Wild 2.0 is a valuable resource for advancing vehicle ReID research and development.
• The DFNet offers a promising approach for handling orientation variations in vehicle ReID.
• The proposed dataset and method facilitate more practical and robust ReID benchmarking.