**MVSGaussian: Fast Generalizable Gaussian Splatting Reconstruction from Multi-View Stereo** **Authors:** Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, Ziwei Liu **Abstract:** MVSGaussian is a novel generalizable 3D Gaussian representation approach derived from Multi-View Stereo (MVS) that efficiently reconstructs unseen scenes. The key contributions are: 1. **MVS-based Geometry-aware Gaussian Representation:** Leverage MVS to encode geometry-aware Gaussian representations and decode them into Gaussian parameters. 2. **Hybrid Gaussian Rendering:** Integrate an efficient volume rendering design for novel view synthesis. 3. **Multi-view Geometric Consistent Aggregation:** Use a consistent aggregation strategy to provide high-quality initialization for per-scene optimization. **Key Challenges:** - **Generalizability:** Overcoming the limitations of existing methods that require per-scene optimization. - **Efficiency:** Achieving real-time rendering with better synthesis quality. - **Optimization:** Designing a fast optimization approach based on the generalizable model. **Contributions:** - **MVSGaussian:** A generalizable Gaussian Splatting method derived from MVS. - **Hybrid Gaussian Rendering:** Enhance generalization with depth-aware volume rendering. - **Consistent Aggregation:** Provide high-quality initialization for fast per-scene optimization. **Experiments:** - **Datasets:** DTU, Real Forward-facing, NeRF Synthetic, Tanks and Temples. - **Results:** MVSGaussian achieves state-of-the-art performance with real-time rendering speed and fast per-scene optimization, outperforming previous methods in terms of synthesis quality and efficiency. **Keywords:** - Generalizable Gaussian Splatting - Multi-View Stereo - Neural Radiance Field - Novel View Synthesis**MVSGaussian: Fast Generalizable Gaussian Splatting Reconstruction from Multi-View Stereo** **Authors:** Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, Ziwei Liu **Abstract:** MVSGaussian is a novel generalizable 3D Gaussian representation approach derived from Multi-View Stereo (MVS) that efficiently reconstructs unseen scenes. The key contributions are: 1. **MVS-based Geometry-aware Gaussian Representation:** Leverage MVS to encode geometry-aware Gaussian representations and decode them into Gaussian parameters. 2. **Hybrid Gaussian Rendering:** Integrate an efficient volume rendering design for novel view synthesis. 3. **Multi-view Geometric Consistent Aggregation:** Use a consistent aggregation strategy to provide high-quality initialization for per-scene optimization. **Key Challenges:** - **Generalizability:** Overcoming the limitations of existing methods that require per-scene optimization. - **Efficiency:** Achieving real-time rendering with better synthesis quality. - **Optimization:** Designing a fast optimization approach based on the generalizable model. **Contributions:** - **MVSGaussian:** A generalizable Gaussian Splatting method derived from MVS. - **Hybrid Gaussian Rendering:** Enhance generalization with depth-aware volume rendering. - **Consistent Aggregation:** Provide high-quality initialization for fast per-scene optimization. **Experiments:** - **Datasets:** DTU, Real Forward-facing, NeRF Synthetic, Tanks and Temples. - **Results:** MVSGaussian achieves state-of-the-art performance with real-time rendering speed and fast per-scene optimization, outperforming previous methods in terms of synthesis quality and efficiency. **Keywords:** - Generalizable Gaussian Splatting - Multi-View Stereo - Neural Radiance Field - Novel View Synthesis