A fully convolutional anchor-free object detector

We propose an improvement on fully convolutional one-stage object detection to improve the overall performance of the model. Firstly, we propose two feature pyramid networks with dense architectures to enhance the semantic features of high levels of feature maps in feature pyramid networks. In addition, we modify the detection head, making it more intuitively to evaluate the prediction results. Finally, the qualities of positive samples in original model are limited by the generated error when the location maps back onto the original image. In this paper, we present an adaptive way to produce positive samples to alleviate this problem by appointing the size of center area adaptively. Besides, some other approaches are adopted to further improve the detector including new loss functions, deformable convolutional modules. We validate our approach on MS COCO validation datasets, and gain a 6.2% AP improvement compared to the original. Based on the backbone of ResNet-50, the detector can achieve an AP of 42.4%.