MAMask: Multi‐feature aggregation instance segmentation with pyramid attention mechanism

Instance segmentation is a fundamental yet challenging vision task. Recently, many instance segmentation methods have attempted to use attention mechanisms to improve model efficiency. However, these methods still ignore the problem of information loss in lateral connection of Feature Pyramid Networks (the supplement operation of low‐resolution, semantically strong features in FPN). The paper presents an effective detection‐based approach named MAMask, which is closely tied to the one‐stage method, Fully Convolutional One‐Stage Object Detection (FCOS). In particular, it adopts the multi‐feature aggregation decoder with pyramid integrate attention (PIA) to instance segmentation. The pyramid integrate attention block can prevent the loss of important channel information by learning richer multi‐scale representation. Meanwhile, it also brings significant improvements in performance for existing FPN‐based frameworks at slight additional computational costs. The proposed MAMask achieves 37.3% in box AP on the COCO dataset. The method outperforms a few recent methods without longer training time. Compared with the current typical algorithms, the proposed method has achieved excellent performance.