A Dataset and Benchmark for 3D Part Recognition from 2D Images

ICLR 2026 Conference Submission13572 Authors

Keywords: 3D Vision, 3D from Single Image, 3D Part recognition

Abstract: While 3D semantic part understanding underpins numerous downstream applications, 3D part detection from images remains underexplored due to limited annotated datasets. To address this, we introduce DST-Part3D, a 3D semantic part dataset with $3,300$ fine-grained 3D part annotations across $475$ shapes from $50$ object categories, paired with $125,000$ realistic synthetic images. DST-Part3D enables training and evaluation of 3D part detection from images, 2D part segmentation via projection, and benchmarking of 3D correspondence quality through transferred part labels. Using this dataset, we develop Part321, an algorithm that recognizes 3D parts in images using only one annotated mesh per category. Part321 establishes mesh-to-mesh and mesh-to-image correspondences to propagate part pseudo-labels across instances, allowing effective 3D part detector training with minimal supervision. Experiments demonstrate that Part321 outperforms previous methods on 3D and 2D part detection tasks. In addition, we use DST-Part3D to analyze the mesh-to-mesh correspondence obtained by different methods leveraging transferred 3d part labels, highlighting the key challenge in 3D part correspondence, which provides insight into future work.

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

Submission Number: 13572