Refining data granularity and feature fusion for boundary refinement in instance segmentation

Abstract

Considerable efforts have been made in the development of current instance segmentation approaches, but the segmentation of mask boundaries remains a challenge. Feature maps with low spatial resolution, along with the small proportion of edge pixels in relation to the total pixel count, lead to inaccurate boundaries in instance masks. Furthermore, the parsing of feature maps in high resolution networks is typically at a low level, making it difficult for the network to learn deeper semantic features. This paper presents improvements to Boundary Patch Refinement (BPR) for Instance Segmentation to address the above issues. First, we improve the bounding box extraction methods utilized in the data processing, refining the granularity of the data. Second, we introduce a feature fusion approach specifically designed to optimize the feature fusion module within the backbone network. Third, we propose Deep enhancement and Memory optimization (DAM), a module that enhances the network’s ability to learn deeper features, improves its efficiency in acquiring semantic information, and substantially reduces the computational overhead during training. Experimental results demonstrate that our network yields notable improvements in both segmentation accuracy and computational efficiency and outperforms existing methods. The code is available at https://github.com/njezmjez/RDGFBR.