FULLY CONVOLUTIONAL NETWORK WITH DENSELY FEATURE FUSION MODELS FOR OBJECT DETECTION

We present Fully Convolutional Networks with Densely Feature Fusion Models (DFF-FCN) which is an effective framework for multi-scale object detection. DFF-FCN reuses the inherent convolutional hierarchical features of regular convolutional neural networks from both forward and backward directions so that the highly semantic features incorporate with shallow high-resolution features complementarily. We propose a new fully convolutional network to build feature pyramids to construct semantic features at all levels using a single neural network so that the multi-scale features all have enough detail and semantic information for object detection tasks. We also add another branch to predict objectness in order to reduce the searching space of objects. Our network runs at the speed of 20 FPS (frame per second) which is faster than Faster R-CNN counterpart and our method gets better detection performance.