NegROI: Click-Centric Uncertainty-Guided Refinement with Scene-Conditioned Negative Prompts for Robust Interactive 3D Segmentation

NegROI is a transformer-based framework for interactive 3D segmentation. It refines a local region around each user click on a finer grid, then fuses the refined logits back into a coarse mask. An uncertainty-driven policy selects ambiguous regions for refinement, while scene-conditioned negative prompts model confusing background structures. The prompts are stabilized with a diversity regularizer, and boundary-aware hard-negative mining emphasizes high-confidence false positives near object boundaries. The abstract reports better click efficiency, fewer false positives, and stronger cross-dataset robustness on ScanNet, S3DIS, and KITTI, but provides no detailed quantitative results.