EvObj: Unsupervised 3D Instance Segmentation via Evolving Object Representations

EvObj introduces an innovative approach to unsupervised 3D instance segmentation, effectively narrowing the geometric domain gap between synthetic pretraining data and actual point clouds. This technique tackles the structural inconsistencies arising from morphological differences and occlusions when transferring object priors from synthetic sources like ShapeNet to real datasets such as ScanNet. It comprises two key components: an object discerning module that continuously refines object candidates for better adaptation to target domains, and an object completion module that reconstructs incomplete geometries post-object identification. Experimental results on both real and synthetic datasets demonstrate its superior performance compared to all baselines, achieving state-of-the-art outcomes. The research is published on arXiv in the computer vision and pattern recognition category.