Residual Joint Attention Network with Graph Structure Inference for Object Detection

Abstract

Most object detectors include three main parts, CNN feature extraction, proposal classification, and duplicate detection removal. In this work, focusing on the improvement of the feature extraction, we propose Residual Joint Attention Network, a convolutional neural network using a residual joint attention module which is composed of a spatial attention branch, a channel attention branch, and a residual learning branch within an advanced object detector with graph structure inference. An attention map generated by the joint attention mechanism is used to weight the original features extracted from a specific layer of VGG16 aiming at performing feature recalibration. Besides, the residual learning mechanism is complementary to the joint attention mechanism and keeps good attributes of the original features. Experimental results show that different branches of our residual joint attention module do not contradict each other. By combining them together, the proposed network obtains higher mAP than many advanced detectors including the baseline on VOC dataset.