DirectPart: Direct 3D part segmentation via multi-view SAM3

This is the authors' abstract. We don't add key points for this paper.

3D part segmentation is critical for applications ranging from robotics to shape modification, but supervised methods require costly point-level annotations and struggle to generalize beyond their training categories. Recent approaches address this limitation by transferring knowledge from 2D base models, pretrained on large-scale image datasets, to 3D shapes through multi-view rendering and backprojection. However, most existing methods still rely on prompt optimization, fine-tuning with few shots, or distillation into dedicated 3D networks. We propose DirectPart, a fully zero-shot pipeline that leverages the Segment Anything Model 3 (SAM3) for 3D part segmentation without any training, fine-tuning, or prompt engineering. The method renders 3D meshes from multiple viewpoints sampled on an icosphere, applies SAM3 with minimal text prompts to segment each view, and aggregates 2D predictions onto 3D surfaces via depth-buffer unprojection and majority voting. We evaluate the full ShapeNetPart benchmark covering all 16 categories. Our method achieves 68.5% mean IoU, 73.6% median IoU, and 83.1% accuracy.