Part detection using object bounding box map attention

Abstract

Part detection is an application of object detection that detects components of objects, it has been drawing attention from industries and academia. However, part detection has relatively low detection accuracy compared to object detection due to small size of part and similarity among parts. To tackle such problems, previous research has utilized object information to improve detection quality, however could only concatenate information of object selected using handcrafted rules. This thesis proposes attention mechanism using OBB map to improve part detection quality. Proposed network consists of 3 parts

1) Object detection network, 2) Attention module using OBB map, 3) Part detection network with attention module. Attention module utilize class and arrangement information to boost important part of feature and suppress unnecessary feature. Experimental results on PASCAL PART dataset showed that proposed network improved detection quality of baseline model.