CVPR2025
AniGrad: Anisotropic Gradient-Adaptive Sampling for 3D Reconstruction From Monocular Video
Noah Stier, Alex Rich, Pradeep Sen, Tobias Höllerer
Abstract
Figure 1 . Reconstruction with adaptive & anisotropic resolution. Our 3D reconstruction system allocates computational resources effectively, using a variable sample pattern that adapts to the local surface complexity and orientation. A) The floor, which is flat and consistently oriented, is accurately reconstructed with a coarse resolution (4 cm). B) The cabinet is coarsely sampled in the vertical dimension because it exhibits no variation along that axis. Horizontally, however, it is sampled finely enough to accurately localize the surface. C) Object boundaries and other high-curvature structures are densely sampled to avoid aliasing, and to maintain high fidelity. The scene shown here is from ScanNet++ [39] .