Weakly supervised 3D instance segmentation without instance-level annotations

Shichao Dong

doi:10.1016/j.patcog.2026.114535

Weakly supervised 3D instance segmentation without instance-level annotations

2026-08-17

0

PRE

AI

DOI：10.1016/j.patcog.2026.114535

Original

Original request for help

Save

Abstract

En 中文

3D instance segmentation often requires instance-level annotations that are costly and scale with the number of objects in a scene. To reduce this annotation burden, we propose a weakly supervised framework for 3D instance segmentation that relies only on categorical semantic labels, without requiring any instance-level supervision. The required semantic annotations can be either dense or extreme sparse (e.g. 0.02% of total points). In the absence of explicit instance cues, we first segment point clouds into primitive fragments and select reliable candidates for learning instance-centric representations. We further construct a recomposed dataset using pseudo instances to facilitate the learning of a multilevel, shape-aware objectness signal. Based on this signal, an asymmetric object inference strategy is applied to handle object cores and boundaries differently, enabling the generation of high-quality pseudo instance labels for iterative training. Experimental results show that the proposed method achieves competitive performance on 3D indoor datasets, such as ScanNet and S3DIS. Moreover, by generating pseudo instance labels from semantic supervision alone, our approach can be used to reduce annotation cost in existing 3D instance segmentation pipelines.

Keywords:

3D scene understanding

Weakly supervised learning

Instance segmentation

Weakly supervised 3D instance segmentation without instance-level annotations

Abstract

Journal

Organization

Cited Papers