Marginal-Aware Framework for 3D Shape Segmentation: Resolving Boundary-Internal Face Imbalance

3D shape segmentation is a fundamental problem in computer vision, supporting applications such as shape reconstruction and semantic understanding. A persistent challenge in learning-based methods is the degradation of performance near part boundaries, commonly attributed to class-level imbalance. In this work, we reveal that the primary source of this issue instead arises from a pronounced imbalance between marginal and internal areas within partitions in 3D shapes, which is largely overlooked by existing methods and leads to systematically poor boundary discrimination. To address this problem, we propose a marginal-aware segmentation framework that explicitly emphasizes boundary localization and relational modeling. The framework is realized by two implementations that serve complementary purposes. Specifically, the staged variant emphasizes interpretability by identifying marginal faces and refining them through topology-aware subgraphs and a Graph Attention Network (GAT), while the end-to-end differentiable integration incorporates our marginal-aware modeling principle into a modern 3D segmentation pipeline (SAMPart3D) by leveraging SAM-derived boundary cues to demonstrate performance gains on state-of-the-art systems. Extensive experiments on PSB, COSEG, and HumanBody demonstrate that the staged variant substantially improves boundary recognition, yielding a 12.32% gain in boundary accuracy over state-of-the-art methods, while the end-to-end integration further improves mIoU from 53.7% to 55.3% on PartObjaverse-Tiny with a modest computational overhead. These results demonstrate that explicit marginal-aware relational modeling is an effective and flexible strategy for improving 3D shape segmentation, particularly in challenging boundary regions.