Unified 3D Instance Segmentation with Contrastive Learning

This paper addresses the challenge of inconsistent 2D instance labels across views in 3D instance segmentation, a problem that arises when extending 2D segmentation to 3D using techniques like 3D Gaussian Splatting and NeRF. The authors propose a unified framework, UniC-Lift, that merges contrastive learning and label consistency steps, improving efficiency and performance. They introduce a learnable feature embedding for segmentation in Gaussian primitives and a novel 'Embedding-to-Label' process. Furthermore, they address object boundary artifacts by incorporating hard-mining techniques, stabilized by a linear layer. The paper's significance lies in its unified approach, improved performance on benchmark datasets, and the novel solutions to boundary artifacts.