MrNeRF (@janusch_patas)
2025-08-01 | โค๏ธ 82 | ๐ 7
NeRF Is a Valuable Assistant for 3D Gaussian Splatting
Contributions: (1) Sharing Mechanism: We first introduce a hash-based network for encoding features in continuous space optimized by NeRF volume rendering. We design strategies to identify potential Gaussian positions. Subsequently, both NeRF and 3DGS share these features to decode additional attributes for their respective spatial points.
(2) Residual Vectors: Due to inherent differences between the NeRF and 3DGS forms, directly using NeRF-optimized features and NeRF-initialized Gaussian positions does not adequately adapt to the 3DGS branch. To address this, we propose explicitly modeling their discrepancies by optimizing residual vectors for both features and positions to personalize and enhance 3DGS performance.
(3) Joint Optimization: We align the attributes and rendering results of spatial points along rays passing through the important Gaussian in the NeRF branch with those in the 3DGS. This reduces feature confusion and ensures mutually beneficial constraints on shared features. Additionally, we leverage NeRFโs continuous spatial query capability to assist in adaptive Gaussian growth, achieving efficient joint optimization across different branches.
๋ฏธ๋์ด
