A-DFPN: Adversarial Learning and Deformation Feature Pyramid Networks for Object Detection

Abstract

In order to weak the variation of the object instance caused by the scale, we innovatively propose an object detection detector based on adversarial learning and deformation feature pyramid: A-DFPN. Firstly, in the feature extraction stage, the concept of Deformation Feature Pyramid Module is proposed. The outstanding advantage is that it can fully extract object features from different convolution layers and objects of different scales. In addition, Two Stage Module is also proposed, it gradually perfects the adjusted anchors in the previous stage through multi-step regression, and locates the position and shape of the object in each RPN to make the location more accurate. At the same time, Mask Module increases the robustness of the detector by spatially blocking certain feature maps or by manipulating feature responses to generate difficult samples. Finally, the final bounding boxes is filtered by soft-NMS. Under the Resnet-101 network architecture, our algorithm achieves the mean average precision of 81.1034% on the Pascal VOC 2007 dataset and 73.52% on the DETRAC dataset, reaching the state-of-the-art detection level.