Atrous Faster R-CNN for Small Scale Object Detection

Deep Convolutional Neural Networks based object detection has made significant progress recent years. However, detecting small scale objects is still a challenging task. This paper addresses the problem and proposes a unified deep neural network building upon the prominent Faster R-CNN framework. This paper has two main contributions. Firstly, an Atrous Region Proposal Network (ARPN) is proposed to explore object contexts at multiple scales by sliding a set of atrous filters with increasing dilation rates over the last convolutional feature map. Secondly, to enrich the representations of small scale image regions, this paper incorporates atrous convolution into Fast R-CNN and proposes a Dense Fast R-CNN (DFRCN), that improves the resolution of the ROI-pooled convolutional feature maps without increasing the number of parameters. In combination of the two, this paper proposes a unified network termed as Atrous Faster R-CNN. On PASCAL object detection challenge dataset, our method achieves superior performance to the state of the arts, especially for small scale objects.