EvObj: Learning Evolving Object-centric Representations for 3D Instance Segmentation without Scene Supervision

2026-05-13

AI Summary

This work addresses the geometric domain shift arising from shape discrepancies and occlusions when applying models pretrained on synthetic data to real-world point clouds, proposing EvObj—a framework for unsupervised 3D instance segmentation. EvObj adaptively captures object priors in the target domain through an object identification module and reconstructs occluded or incomplete local geometries via an object completion module. It further integrates object-centric representation learning with a dynamic proposal refinement mechanism to jointly optimize instance segmentation and geometric reconstruction without requiring scene-level annotations. Extensive experiments demonstrate that EvObj significantly outperforms existing unsupervised methods across multiple real and synthetic datasets, achieving state-of-the-art performance.

Abstract

We introduce EvObj for unsupervised 3D instance segmentation that bridges the geometric domain gap between synthetic pretraining data and real-world point clouds. Current methods suffer from structural discrepancies when transferring object priors from synthetic datasets (e.g., ShapeNet) to real scans (e.g., ScanNet), particularly due to morphological variations and occlusion artifacts. To address this, EvObj integrates two innovative modules: (1) An object discerning module that dynamically refines object candidates, enabling continuous adaptation of object priors to target domains; and (2) An object completion module that reconstructs partial geometries after discovering objects. We conduct extensive experiments on both real-world and synthetic datasets, demonstrating superior 3D object segmentation performance over all baselines while achieving state-of-the-art results.

Problem

• 3D instance segmentation
• domain gap
• unsupervised learning
• object priors
• point clouds

Innovation

• unsupervised 3D instance segmentation
• object-centric representation
• domain adaptation
• object completion
• point cloud