Densely convolutional and feature fused object detector

In this paper, we propose a novel deep convolutional network for object detection named densely convolutional and feature fused object detector(DCFF-Net), which is a one-stage object detector from scratch similarly to DSOD. The base network is stacking by several densely convolutional blocks to extract the powerful semantic information, and the feature fusion module is used to obtain the enriching features by fusing the extracted feature maps from different convolutional layers. In the fusion module, the feature maps are concatenated of three adjacent scales, which are from the features extracted by the convolution with big kernels, the features extracted by down-sampling pooling and the features extracted by up-sampling deconvolution. The fused feature pyramid has more representative information and gets better performances when it is fed to the final multibox detectors. On the Pascal VOC 2007/2012 and MS COCO, our network achieves better results than DSOD and several methods with pre-training models. The experimental results show that our proposed network has better detection performance by the aid of the fusion of different layers’ feature maps, especially on small objects and occluded objects.