Stable Surface Regularization for Fast Few-Shot NeRF

TL;DR

This work addresses the challenge of fast, high-quality novel-view synthesis from few input views by introducing Annealing Signed Distance Function (ASDF) loss, a stable, coarse-to-fine surface regularization that overcomes the limitations of the traditional Eikonal loss in sparse data regimes. By combining multi-level voxel grids for SDF and color with monocular priors and a jointly optimized ASDF term, the approach delivers robust geometry and appearance reconstruction while achieving substantial training-time reductions (up to 30–45x) compared to prior NeRF methods. The method is validated on ScanNet and NeRF-Real, showing competitive visual and depth metrics and improved stability; ablations confirm the benefits of ASDF and geometric priors. Overall, the technique offers a practical path to real-world, fast few-shot NeRF applications with strong geometric regularization under sparse supervision.

Abstract

This paper proposes an algorithm for synthesizing novel views under few-shot setup. The main concept is to develop a stable surface regularization technique called Annealing Signed Distance Function (ASDF), which anneals the surface in a coarse-to-fine manner to accelerate convergence speed. We observe that the Eikonal loss - which is a widely known geometric regularization - requires dense training signal to shape different level-sets of SDF, leading to low-fidelity results under few-shot training. In contrast, the proposed surface regularization successfully reconstructs scenes and produce high-fidelity geometry with stable training. Our method is further accelerated by utilizing grid representation and monocular geometric priors. Finally, the proposed approach is up to 45 times faster than existing few-shot novel view synthesis methods, and it produces comparable results in the ScanNet dataset and NeRF-Real dataset.

Figures (10)

• Figure 1: Our method can synthesize novel views within 30 minutes by utilizing multi-level voxel grid optimization. To overcome the limitation of novel view synthesis with sparse input images, we utilized additional strong geometric cues and our novel geometric smoothing loss, Annealing SDF loss.
• Figure 2: Novel view synthesis comparisons with state-of-the-art works (a) DSNeRF deng2022depth (b) DDPNeRF roessle2022dense (c) voxel grids based approach sun2022direct with eikonal loss gropp2020implicit and depth supervision, and (d) ours. Top row: color images, bottom row: depth maps.
• Figure 3: The overall pipeline. The proposed architecture utilizes structure from motion to extract sparse 3D information and camera poses from sparse input views, while off-the-shelf depth and surface normal are obtained from a pretrained network eftekhar2021omnidata. Points are sampled with respect to the camera from the structure from motion, and the feature value at corresponding points is extracted. SDF values, gradient of SDF values, and RGB values are decoded by simple MLP decoder, and RGB, surface normal, and depth values are extracted along the ray using volumetric rendering. The rendered RGB, depth map, and surface normal are supervised with their respective label using the loss functions $L_C, L_D$, and $L_N$, while SDF values are supervised with the loss functions $L_{\text{ASDF}}$ which is composed by $L_{\text{GS}}$, and $L_{\text{wEik}}$.
• Figure 4: Qualitative results in a real-world scene. It shows that the Eikonal loss has difficulties in reconstructing surface geometry from a few training images which results in over-smooth depth and color rendering qualities. Top : images, bottom : depth maps.
• Figure 5: The visualization for Annealing SDF loss. The truncated bounds $b$ decrease from $b_\text{max}$, camera origin $o$, to $b_\text{opt}$ by following iteration process to capture detailed geometry by reducing smoothing area. $L_\text{GS}$ is applied to $\textbf{p}(t)$ within grey region $R'$.