CDSF: A curvature-driven semi-supervised framework with dynamic receptive fields for fine-grained vehicle component segmentation

Zhili Gong, Chunyuan Zheng, Shanshan Huang, Huayi Yang, Guoxin Su, Li Liu

Abstract: Highlights•Introduce a curvature-aware semi-supervised framework (CSDF) that boosts vehicle exterior segmentation with scarce labels.•Propose a Dynamic Adaptation Module that adjusts receptive fields and confidence thresholds via local curvature, recovering fine details.•Present an Edge-Guided Aggregation Module that aligns feature sampling with boundary normals to suppress cross-part leakage.•Release a high-quality 18-class vehicle exterior dataset and show consistent mIoU gains over SOTA methods on private and public benchmarks.

External IDs:dblp:journals/kbs/GongZHYSL25