Multi-Query Vehicle Re-Identification: Viewpoint-Conditioned Network, Unified Dataset and New Metric

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current vehicle re-identification (Re-ID) methods struggle with limited information from single queries, hindering performance in complex surveillance networks.
• Single-query Re-ID faces significant limitations due to viewpoint variations and incomplete data.

Purpose of the Study:

• To introduce and address the task of multi-query vehicle Re-ID, leveraging multiple queries to overcome single-query limitations.
• To enhance vehicle representation and recognition accuracy in large-scale transportation surveillance systems.

Main Methods:

• A novel viewpoint-conditioned network (VCNet) is proposed to adaptively integrate complementary information from diverse vehicle viewpoints.
• A cross-view feature recovery module is developed to address missing vehicle viewpoints by learning feature correlations.
• A unified benchmark dataset from a real-life transportation surveillance system (6142 cameras) is created for multi-query Re-ID evaluation.
• A new evaluation metric, mean cross-scene precision (mCSP), is designed to assess cross-scene recognition capabilities.

Main Results:

• The proposed VCNet method demonstrates superior performance compared to existing methods in comprehensive experiments.
• The effectiveness of the mCSP metric in evaluating multi-query vehicle Re-ID is validated.
• The developed dataset provides a robust benchmark for future research in this domain.

Conclusions:

• Multi-query vehicle Re-ID offers a more realistic and effective approach for surveillance applications.
• The VCNet architecture and mCSP metric significantly advance the state-of-the-art in vehicle Re-ID.
• The study provides valuable resources (code and dataset) for the research community.