SAM-HQ and Prior-Guided Label Reassignment Boost 3D Gaussian Splatting Segmentation
A team of researchers has unveiled a groundbreaking technique in 3D Gaussian Splatting (3D-GS), aimed at refining object segmentation for applications such as object removal and recoloring. This innovative method, explained in a preprint on arXiv, employs the Segment Anything Model High Quality (SAM-HQ) to create detailed 2D masks, addressing earlier challenges in boundary precision. Additionally, it introduces a prior-guided label reassignment approach for improved 3D object segmentation, ensuring visual consistency across various perspectives. This advancement promises high segmentation fidelity and allows for interactive, real-time editing, significantly enhancing visual output quality. The preprint was published on May 23, 2025.
Key facts
- Framework uses SAM-HQ for accurate 2D masks.
- Prior-guided label reassignment enforces multiview consistency.
- Enables real-time editing of 3D scenes.
- Achieves state-of-the-art segmentation accuracy.
- Addresses view inconsistencies and coarse masks in prior methods.
- Supports object removal, extraction, and recoloring.
- Maintains high visual fidelity.
- Preprint available on arXiv (2605.16065).
Entities
Institutions
- arXiv