Reverse Densely Connected Feature Pyramid Network for Object Detection

The wide and extreme diversity of object size is an everlasting challenging issue in object detection research. To address this problem, we propose Reverse Densely Connected Feature Pyramid Network (Rev-Dense FPN), a novel multi-scale feature transformation and fusion method for object detection. Through reverse dense connection, we directly fuse all the feature maps of higher levels than the current one. This avoids useful contextual information on the higher level to vanish when passed down to lower levels, which is a key disadvantage of widely used feature fusion paradigms such as recursive top-down connection. Therefore, a more powerful hierarchical representation structure can be obtained by effectively aggregating multi-level contexts. We apply Rev-Dense FPN on SSD framework, which reaches 81.1% mAP (mean average precision) on the PASCAL VOC 2007 dataset and 31.2 AP on the MS COCO dataset. The results show that Rev-Dense FPN is more effective in dealing with diversified object sizes.