Vehicle re-identification using machine learning

Abstract

Vehicle Re-ID aims to retrieve images of the same vehicle across non-overlapping cameras. The key challenges lie in the subtle inter-class discrepancy caused by near-duplicated identities and the significant intra-class distance due to diverse factors, including illumination, viewpoints, and background interference. This thesis starts with reviewing the development history of vehicle Re-ID and proposes a Part-Mentored Attention Network (PMANet) consisting of a Part Attention Network (PANet) for weakly-supervised vehicle part localization and a Part-Mentored Network (PMNet) for mentoring the global and local feature aggregation. Firstly, PANet predicts a foreground mask and pinpoints K prominent vehicle parts without additional part-level supervision. Secondly, PMNet applies multi-scale soft attention on localized regions and compensates inaccurate part masks with part-guided learning. PANet and PMNet construct a two-stage attention structure to perform a coarse-to-fine search among identities. Finally, we address this Re-ID issue as a multi-task problem and employ Homoscedastic Uncertainty Learning to automatically balance the loss weightings. Experimental results show that our approach outperforms recent state-of-the-art methods by averagely 2.63% in CMC@1 on VehicleID and 2.2% in mAP on VeRi776.