Feature Aggregation Network with Tri-Hybrid Loss for Instance Segmentation
Zeping Zhou, Yongxiong Wang, Jin Peng · 2021
Limited by the size of feature maps in the mask head, the operation of simply stacking convolutional blocks in previous instance segmentation methods cannot obtain comprehensive features. In contrast, we design a Feature Aggregation Net-work (FAN) where three separate modules are respectively responsible for extracting salient features from the dimensions of channel, space, and scale. Then, these salient features are further aggregated based on their affinities generated by an Affinity Computation Module (ACM) with the original features. In this way, features become more distinct, which is conducive to the subsequent mask prediction. To further improve the segmentation performance of hard examples with-out introducing extra inference overhead, we also propose a novel loss function named Tri-hybrid loss where the optimization can be simultaneously performed from the perspectives of pixel, boundary, and instance. With these improvements, our model outperforms state-of-the-arts with 39.5 mask AP on COCO test-dev2017.