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

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Computer Vision · Depth: Expert, long

Summary

NegROI is a novel transformer-based interactive framework designed for robust 3D segmentation in point clouds, addressing issues like coarse voxel resolution blurring fine boundaries and hard false positives from confusing background structures. It integrates click-centric multi-resolution refinement with scene-conditioned negative prompts. The system refines a local Region Of Interest (ROI) around the current click on a finer grid (fine scale η=2) and fuses these refined logits back to the coarse mask. It employs uncertainty-driven selective refinement (threshold τ=0.20), models background patterns via K=8 scene-conditioned negative prompts, and uses boundary-aware hard negative mining (k=256 voxels) with a diversity regularizer. Experiments on ScanNet, S3DIS, and KITTI demonstrate improved click efficiency, reduced false positives, and stronger cross-dataset robustness compared to baselines.

Key takeaway

For Machine Learning Engineers developing interactive 3D segmentation systems, NegROI offers a robust approach to improve boundary accuracy and reduce false positives. You should consider integrating click-centric multi-resolution ROI refinement and scene-conditioned negative prompts into your models. This strategy enhances cross-dataset generalization and click efficiency, especially for sparse interactions, by focusing computation on ambiguous regions and explicitly modeling background distractors.

Key insights

NegROI enhances interactive 3D segmentation by combining local, fine-grid refinement with explicit background modeling using scene-conditioned negative prompts.

Principles

• Refine only ambiguous regions for efficiency.
• Model background distractors explicitly.
• Regularize prompt diversity to prevent collapse.

Method

NegROI uses a sparse voxel backbone and two-way transformer. It decodes coarse logits, then refines a click-centric ROI on a finer grid, fusing results. Scene-conditioned negative prompts are appended to interaction tokens, supervised by boundary-aware hard negative mining and diversity regularization.

In practice

• Implement multi-resolution refinement in ROIs.
• Use negative prompts to suppress background.
• Apply uncertainty-driven selection for efficiency.

Topics

• Interactive 3D Segmentation
• Point Cloud Segmentation
• Transformer Decoders
• Negative Prompts
• Region of Interest (ROI) Refinement
• Cross-Dataset Robustness

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• NegROI: Click-Centric Uncertainty-Guided Refinement with Scene-Conditioned Negative Prompts for Robust Interactive 3D Segmentation — NegROI refines 3D segmentation using click-centric multi-resolution processing and scene-conditioned negative prompts for robustness across diverse datasets.
• NG-GS: NeRF-Guided 3D Gaussian Splatting Segmentation — NG-GS enhances 3D Gaussian Splatting segmentation by integrating NeRF-guided continuous feature fields to resolve boundary artifacts.
• Texture-Shape Bias Balancing for Robust Synthetic-to-Real Semantic Segmentation in Automotive NIR Imagery — A generative augmentation framework reduces the synthetic-to-real domain gap for automotive NIR semantic segmentation by balancing texture-shape bias.
• HOI-aware Adaptive Network for Weakly-supervised Action Segmentation — Leveraging human-object interaction (HOI) context dynamically improves weakly-supervised action segmentation accuracy for ambiguous actions.
• Boosting Zero-Shot 3D Style Transfer with 2D Pre-trained Priors — Integrating 2D pre-trained decoders into 3D style transfer pipelines effectively mitigates data scarcity and enhances stylization quality.
• Segmentation before Answering: Pixel Grounding for MLLM Visual Reasoning — Pixel-level segmentation significantly enhances MLLM visual reasoning by precisely isolating regions of interest, outperforming bounding box methods.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.