Faster R-CNN with Attention Feature Map for Robust Object Detection

Abstract

This paper describes the improved object detection method from Faster R-CNN using an attention feature map in RPN. The research work adopts Faster R-CNN is used as the input feature map of the RPN using the last block in the backbone network, but the proposed method takes the created feature maps from the combination of dilated convolution and attention feature map for RPN networks. Attention feature map takes the high probability to object and emphasize the location of object. With the new feature map, the proposed bounding box and class determine the final output for object detection. As backbone network, ResNet50, ResNet101 and ResNet152 have trained on ImageNet. On PASCAL VOC 2007, our proposed method achieves 79.83% mAP that is 73.2% mAP.