Multi-scale Vertical Cross-layer Feature Aggregation and Attention Fusion Network for Object Detection

Scale imbalance is one of the primary limitations for object detection. To tackle such a problem, existing methods such as FPN usually integrate the features at different scales, which suffers from the inconsistence of different high-level and low-level features due to the straightforward combination. In this paper, we propose a multi-scale vertical cross-layer feature aggregation and attention fusion network which not only has bottom-up and top-down pathways with lateral connections, but also adds cross-layer paths in the vertical direction. The proposed method can boost information flow and shorten the information path between high-level and low-level features. An attention fusion module is also introduced to obtain the internal correlation between local, global and contextual information of other feature layers. In order to optimize the anchor configurations, a differential evolution algorithm is employed to reconfigure the ratios and scales of anchors. Experimental results show that the proposed method achieves superior detection performance on the public dataset PASCAL VOC.