Multi-Scale Residual Aggregation Feature Pyramid Network for Object Detection

The effective use of multi-scale features remains an open problem for object detection tasks. Recently, proposed detectors have usually used Feature Pyramid Networks (FPN) to fuse features. Since a relatively simple feature map fusion approach, it can lead the loss or misalignment semantic information in process. Several works demonstrated that using bottom-up structure Network shorten path between lower layers and topmost feature, allowing adequate exchange from different layers. We further enhance by proposing residual aggregation (MSRA-FPN), which uses unidirectional cross-layer module aggregate multiple triangular layer. In addition, we introduce Residual Squeeze Excitation Module mitigate aliasing effects occur when are aggregated. MSRA-FPN enhances high-level maps, mitigates decay during fusion, capability model large objects. It is experimentally our improves performance three baseline models 0.5–1.9% on PASCAL VOC dataset also quite competitive with other state-of-the-art FPN methods. On MS COCO dataset, method improve 0.8% model’s 1.8%. To validate effectiveness detection, constructed Thangka Figure Dataset conducted comparative experiments. 2.9–4.7% this reach up 71.2%.