Contour-Based Instance Segmentation With the Aid of Pixel-Wise Classification and Topological Regularization

Abstract

Contour-based instance segmentation offers several advantages, such as suppressing isolated outlines and enabling efficient inference, by directly producing vectorized polygons that are well suited for precision manufacturing (e.g., semiconductor/PCB inspection) and geospatial applications (e.g., GIS building extraction). However, existing contour-evolution frameworks remain challenged by two practical issues: 1) limited spatial context when refinement relies on features sampled only at sparse vertices, which hinders the accurate recovery of rapidly varying boundaries, and 2) topological instability, where iterative updates frequently yield self-intersecting (non-simple) polygons that are invalid for downstream metrology and geometry processing. To address these limitations, we propose a contour-based framework with two key contributions. First, we introduce a dense-score-guided refinement scheme that augments contour evolution with an auxiliary pixel classification head predicting a dense interior-exterior score map; this grid-level signal is fused with backbone features and used to guide vertex updates, improving boundary localization. Second, we propose an energy-inspired topological loss that introduces explicit, differentiable geometric constraints on total contour length and curvature, effectively reducing self-intersections during training. Extensive experiments on precision manufacturing datasets (PCB bond finger dataset) and remote-sensing benchmarks (WHU building dataset) demonstrate that the proposed approach consistently improves boundary accuracy and significantly increases the proportion of geometrically valid, simple (i.e., non-self-intersecting) polygons compared with state-of-the-art contour-based instance segmentation methods.