Gaussian Prompter: Linking 2D Prompts for 3D Gaussian Segmentation.

Area of Science:

• Computer Vision
• 3D Scene Understanding
• Computational Geometry

Background:

• Interactive 3D segmentation is vital for scene understanding and manipulation.
• Current methods lack volumetric completeness and segmentation accuracy due to insufficient integration of multi-view 2D prompt-based segmentations.

Purpose of the Study:

• To introduce Gaussian Prompter, a novel approach for 3D segmentation within 3D Gaussian Splatting.
• To improve volumetric completeness and segmentation accuracy by linking multi-view 2D prompts.

Main Methods:

• Gaussian Prompter integrates a Gaussian-centric segmentation paradigm.
• GaussBlend aggregates multi-view 2D masks into a cohesive 3D segmentation.
• PinPrompt utilizes high-confidence adjacent view prompts to enhance precision.
• Introduced SegMip-360 dataset with over 350 annotated masks across seven scenes.

Main Results:

• Gaussian Prompter significantly outperforms existing state-of-the-art methods.
• Demonstrated superior performance in both segmentation accuracy and completeness.
• The SegMip-360 dataset provides a valuable resource for 3D segmentation research.

Conclusions:

• Gaussian Prompter effectively addresses limitations in current 3D segmentation techniques.
• The proposed method offers a robust solution for accurate and complete 3D segmentation.
• Future research can build upon this framework and dataset for further advancements.