Mask R-CNN with Feature Pyramid Attention for Instance Segmentation

Abstract

In this paper, a new network for instance semantic segmentation is proposed, which is combined with an attention mechanism and is theoretically effective and simple. In order to perform spatial pyramid attention on high-level output, the proposed network extends Mask R-CNN by exploiting a Feature Pyramid Attention module. The proposed network is simple to be trained and extended. Moreover, the new network is easy to be generalized to other tasks. For example, the new network is suitable for bounding-box detection, person keypoint detection and instance segmentation in the same framework. Compared with the other methods, including MNC, FCIS and Mask R-CNN, which are the champion of the COCO 2015-2017 semantic segmentation challenges respectively, the newly proposed network is superior to them on small scale, and the APs of the result is 14.7. On other scales and for AR metric, the proposed network achieves the comparable performance with the state-of-the-art.