Representative view selection and hard view learning for multiview 3D model classification

Abstract

View-based 3D model classification benefits from mature 2D visual backbones, but dense multi-view rendering usually contains many repeated observations. This paper therefore focuses on two practical questions: how to keep a compact set of representative views, and how to make the classifier learn more from difficult views rather than only from visually obvious ones. Each 3D model is first rendered into 80 views. A three-stage selector composed of farthest point sampling, submodular coverage optimization, and determinantal point process refinement is then used to retain 20 representative views. Four low-confidence views are further mined from this subset and reused as hard samples during training. The final prediction is obtained by confidence-weighted voting, and a hybrid NLL-focal loss strengthens ambiguous views. Experiments on ModelNet40 show that the proposed framework achieves 90.60% accuracy, indicating that representative selection and hard-view reinforcement provide an effective balance between accuracy, compactness, and robustness.