MrNeRF (@janusch_patas)
2025-02-28 | โค๏ธ 178 | ๐ 21
No Parameters, No Problem: 3D Gaussian Splatting without Camera Intrinsics and Extrinsics
Contributions: โข We theoretically derive the gradients of the focal length in relation to 3DGS training to update the camera intrinsics, thereby eliminating any prior information about the camera parameters for 3DGS training.
โข To our knowledge, for the first time, we introduce a joint optimization approach for camera parameters and 3DGS. We achieve this by initializing a set of 3D Gaussians and enforcing trajectory and scale constraints, which allows us to apply multi-view consistency and reprojection loss to estimate camera parameters, resulting in a robust 3DGS.
โข On both public and synthetic datasets, our approach outperforms previous methods that require camera intrinsics and achieves SOTA performance on novel view synthesis.
๋ฏธ๋์ด
๐ Related
Auto-generated - needs manual review