DualHead for One-stage Object Detection Networks with Receptive Field Enhancement

The ordinary detection head has a simple structure in one-stage object detection networks, leading to its receptive field being too small to completely cover the feature region of some objects with a large aspect ratio. Furthermore, the detection precision of networks is also reduced. To solve this problem, we propose a dual detection head, called DualHead, to enhance the receptive field and improve the detection precision. The DualHead is composed of two parallel sub-heads with different receptive fields: the sub-head with a small receptive field extracts dense small range features, and the other with a large receptive field extracts sparse large range features. By fusing the feature maps output by two sub-heads with the proposed channel-wise reorganization convolution fusion (CRCF) module, the receptive field of DualHead is about 4.5 times larger than that of the ordinary head, so that it is enough to cover the whole feature region of all objects to be detected. The experiments on the MSCOCO 2017 dataset show that DualHead improves the detection precision AP by 1.2% and 0.9% of ATSS with ResNet-50 and Swin-T as the backbone.