Real-Time Vehicle Orientation Classification and Viewpoint-Aware Vehicle Re-Identification

Vehicle re-identification (re-ID) is based on identity matching of vehicles across non-overlapping camera views. Recently, the research vehicle re-ID attracts increased attention, mainly due to its prominent industrial applications, such as post-crime analysis, traffic flow and wide-area tracking. However, despite interest, problem remains be challenging. One most significant difficulties large viewpoint variations non-standardized placements. In this study, improve robustness against while preserving algorithm efficiency, we exploit use orientation information. First, analyze benchmark various deep learning architectures in terms performance, memory use, cost applicability classification. Secondly, extracted information utilized task. For this, propose a viewpoint-aware multi-branch network that improves performance without increasing forward inference time. Third, introduce mini-batching approach which yields improved training higher performance. The experiments show an increase 4.0% mAP 4.4% rank-1 score popular VeRi dataset with proposed strategy, overall, 2.2% 3.8% compared ResNet-50 baseline .