Generalizable 3D Edge Detection For Soft & Hard Features

ICLR 2026 Conference Submission14030 Authors

Keywords: geometric learning, 3D reconstruction, edge detection, segmentation

TL;DR: Being able to split 3D shapes into fine-grain geometric pieces based on geometry alone, not absolute labels, is important for understanding shapes. We present an improved method for identifying and partititioning edges within shapes.

Abstract: Understanding 3D objects based on their geometric and physical properties--independent of predefined labels--is essential for creating, modifying, and using the objects in diverse contexts. However, most machine learning approaches in the 3D domain rely heavily on semantic or primitive labeled-data to achieve these tasks. We present a 3D edge detection algorithm that decomposes point clouds into precise geometric components without relying on primitives or semantic labels. This enables us to tackle datasets of freeform, entirely unrestricted objects (as in the Thang3D dataset) that are challenging, and in many cases impossible, for current models in the literature to segment, reconstruct, or produce parametrically. Additionally, we achieve state-of-the-art (SOTA) edge detection accuracy on both the complex Fusion360 Segmentation, Thang3D, and simpler standard ABC benchmarks. Our approach maintains reliable edge detection on soft features where most existing models fail. In addition, when the detected edges are used as input for segmentation, our method outperforms recent segmentation models on intricate geometries. This framework provides a robust and generalizable foundation for edge-aware analysis, segmentation, and generation of diverse 3D shapes well beyond what can be easily labeled by humans.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 14030