Problem
Research questions and friction points this paper is trying to address.
feed-forward 3D Gaussian Splatting
pixel-aligned formulation
3D geometric priors
scene-level reconstruction
Gaussian representation
AnchorSplat: Feed-Forward 3D Gaussian SplattingWith 3D Geometric Priors
🤖 AI Summary
Existing feedforward Gaussian reconstruction methods rely on pixel-wise alignment, which is constrained by image resolution and the number of input views, hindering efficient and high-fidelity scene-level reconstruction. This work proposes AnchorSplat, a feedforward 3D Gaussian splatting framework grounded in 3D geometric priors. By modeling scenes through anchor-based alignment in 3D space, AnchorSplat eliminates dependence on pixel coordinates and incorporates a Gaussian refiner to enhance representation quality. The approach substantially reduces the required number of Gaussian primitives while improving reconstruction consistency and computational efficiency. On the ScanNet++ v2 NVS benchmark, AnchorSplat achieves state-of-the-art view synthesis quality and reconstruction fidelity using fewer Gaussians than existing methods.
📝 Abstract
Recent feed-forward Gaussian reconstruction models adopt a pixel-aligned formulation that maps each 2D pixel to a 3D Gaussian, entangling Gaussian representations tightly with the input images. In this paper, we propose AnchorSplat, a novel feed-forward 3DGS framework for scene-level reconstruction that represents the scene directly in 3D space. AnchorSplat introduces an anchor-aligned Gaussian representation guided by 3D geometric priors (e.g., sparse point clouds, voxels, or RGB-D point clouds), enabling a more geometry-aware renderable 3D Gaussians that is independent of image resolution and number of views. This design substantially reduces the number of required Gaussians, improving computational efficiency while enhancing reconstruction fidelity. Beyond the anchor-aligned design, we utilize a Gaussian Refiner to adjust the intermediate Gaussiansy via merely a few forward passes. Experiments on the ScanNet++ v2 NVS benchmark demonstrate the SOTA performance, outperforming previous methods with more view-consistent and substantially fewer Gaussian primitives.
Innovation
Methods, ideas, or system contributions that make the work stand out.
anchor-aligned representation
3D geometric priors
feed-forward 3D Gaussian Splatting
Gaussian Refiner
scene-level reconstruction
🔎 Similar Papers
No similar papers found.
