Feature Combination Based on Receptive Fields and Cross-Fusion Feature Pyramid for Object Detection

Abstract

Currently, the state-of-the-art method about object detector in image mainly depends on deep backbones, such as ResNet-50, DarkNet-53, ResNet-101 and DenseNet-169, which benefits for their powerful capability of feature representations but suffers from high computational cost. On the basis of fast lightweight backbone network (i.e., VGG-16), this paper improves the capability of feature representations by combining features of different receptive fields and cross-fusing feature pyramids, and finally establishes a fast and accurate detector. The architecture of our model is designed to integrate FC-CF Net with two sub-modules: FC module and CF module. Inspired by the structure of receptive fields in visual systems of human, we propose a novel method about Feature Combination Based on Receptive Fields module (FC module), which takes the relationship between the size and eccentricity of receptive fields into account, and then combine them with original features for increasing the receptive field and information of the feature map. Furthermore, based on the structure of FPN (Feature Pyramid Network), we design a novel Cross-Fusion Feature Pyramid module (CF module), which combines top-down and bottom-up connections to fuse features across scales, and achieves high-level semantic feature map at all scales. Extensive experiments on PASCAL VOC 2007 and 2012 demonstrate that FC-CF Net achieves state-of-the-art detection accuracy (i.e. 82.4% mAP, 80.5% mAP) with high efficiency (i.e. 69 FPS, 35 FPS).