BFF R-CNN: Balanced Feature Fusion for Object Detection

In the neck part of a two-stage object detection network, feature fusion is generally carried out in either top-down or bottom-up manner. However, two types imbalance may exist: model and gradient region interest extraction layer due to scale changes objects. The deeper network is, more abstract learned features are, that say, semantic information can be extracted. extracted image background, spatial location, other resolution are less. contrast, shallow learn little information, but lot location information. We propose Both Ends Centre Multiple Layers (BEtM) method solve problem Multi-level Region Interest Feature Extraction (MRoIE) problem. combination with Region-based Convolutional Neural Network (R-CNN) framework, our Balanced Fusion (BFF) offers significantly improved performance compared Faster R-CNN architecture. On MS COCO 2017 dataset, it achieves an average precision (AP) 1.9 points 3.2 higher than those Pyramid (FPN) framework Generic Extractor (GRoIE) respectively.