Multi-Layer Features Based Self-Support Network for Few-Shot Segmentation

Few-shot semantic segmentation aims to make the network identify the object regions of new categories with only a few annotated samples. The key to the few-shot segmentation is to complete information interaction between support and query images, and avoid overfitting as much as possible. For combining the advantages of metric-based prototype learning and visual correspondence, we propose a Multi-Layer Features Based Self-Support Network (MFSNet). In MFSNet, we divide the interactive part of support and query features into two stages, multi-layer similarity feature stage and self-support prototype fusion stage. The multi-layer similarity feature stage focuses more on extracting the details of the segmented object, and the self-support prototype fusion stage refine the general feature information of the object. The two stages are linked together through self-support strategies. In addition, we propose a Fusion Block in MFSNet to establish a more robust semantic relationship between multi-layer features. MFSNet is tested on two benchmark semantic segmentation datasets, Pascal-5
ⁱ
and COCO-20
ⁱ
, and achieves new state-of-the-art performances on 1-shot segmentation tasks.