Part-Guided Attention Learning for Vehicle Instance Retrieval

Vehicle instance retrieval (IR) often requires one to recognize the fine-grained visual differences between vehicles. Besides holistic appearance of vehicles which is easily affected by viewpoint variation and distortion, vehicle parts also provide crucial cues differentiate near-identical Motivated these observations, we introduce a Part-Guided Attention Network (PGAN) pinpoint prominent part regions effectively combine global local information for discriminative feature learning. PGAN first detects locations different components salient regardless identity, serves as bottom-up attention narrow down possible searching regions. To estimate importance detected parts, propose Part Module (PAM) adaptively locate most with high-attention weights suppress distraction irrelevant relatively low weights. The PAM guided identification loss therefore provides top-down that enables attention be calculated at level car other Finally, aggregate features together improve performance further. combines part-guided bottom-up top-down attention, in an end-to-end framework. Extensive experiments demonstrate proposed method achieves new state-of-the-art IR on four large-scale benchmark datasets.1