HYPER FEATURE FUSION PYRAMID NETWORK FOR OBJECT DETECTION

We present Hyper Feature Fusion Pyramid Network (HFFP-net), an efficient framework for object detection which firstly fuses features from multiple layers and then builds a new branch to construct feature pyramids based on the fused feature. HFFPnet reuses the hierarchical features from multiple layers of convolutional neural networks to build feature pyramids to construct semantic features at all levels so that the multi-scale features of different layers are highly enriched. We also add another branch in the detection head to predict objectness to reduce easy negative candidates. We propose several different feature fusion models and we have done several experiments to show the advantages of the proposed approach. Our network runs at the speed of 20 FPS (frame per second) which is faster than Faster R-CNN counterpart and our method achieves competitive detection performance.