The article introduces ‘Compact and Swift Segmenting 3D Gaussians (CoSSegGaussians)’, a method designed to perform 3D scene segmentation swiftly and compactly using only RGB images as input. It improves upon previous NeRF-based segmentation methods that are time-consuming due to their reliance on implicit or voxel neural scene representation and ray-marching volume rendering. The CoSSegGaussians model optimizes Gaussian points position, convariance, and color attributes, distills multi-scale DINO features, and incorporates them into a shallow decoding network to achieve high-quality zero-shot scene segmentation. The model outperforms other segmentation methods in terms of speed and quality.
Publication date: 11 Jan 2024
Project Page: https://David-Dou.github.io/CoSSegGaussians
Paper: https://arxiv.org/pdf/2401.05925
