3D Surface Reconstruction with Enhanced High-Frequency Details

TL;DR

FreNeuS addresses the tendency of neural implicit 3D reconstruction to produce overly smooth surfaces by explicitly exploiting high-frequency information. It detects high-frequency image regions via gradient-based analysis, dynamically biases ray sampling toward detail-rich areas, and applies a pixel-space constraint to reinforce high-frequency details during optimization, all within a NeuS-compatible framework. The approach yields improved surface fidelity and texture detail on DTU and NeRF-synthetic datasets, with ablations confirming the contributions of high-frequency detection, sampling, and constraints. The method generalizes to other NeuS-based models, offering a practical, low-overhead enhancement for local detail reconstruction in neural implicit surfaces.

Abstract

Neural implicit 3D reconstruction can reproduce shapes without 3D supervision, and it learns the 3D scene through volume rendering methods and neural implicit representations. Current neural surface reconstruction methods tend to randomly sample the entire image, making it difficult to learn high-frequency details on the surface, and thus the reconstruction results tend to be too smooth. We designed a method (FreNeuS) based on high-frequency information to solve the problem of insufficient surface detail. Specifically, FreNeuS uses pixel gradient changes to easily acquire high-frequency regions in an image and uses the obtained high-frequency information to guide surface detail reconstruction. High-frequency information is first used to guide the dynamic sampling of rays, applying different sampling strategies according to variations in high-frequency regions. To further enhance the focus on surface details, we have designed a high-frequency weighting method that constrains the representation of high-frequency details during the reconstruction process. Qualitative and quantitative results show that our method can reconstruct fine surface details and obtain better surface reconstruction quality compared to existing methods. In addition, our method is more applicable and can be generalized to any NeuS-based work.

Figures (6)

• Figure 1: Overview of our proposed Fre-NeuS. We first denoise the input RGB image using Gaussian filtering, then obtain the frequency distribution through the gradient detection module. Based on this distribution, we dynamically adjust the sampling rate to ensure sufficient capture of high-frequency details. We utilize the obtained high-frequency map to create a weight map in pixel space and adjust the optimization of pixels at different frequencies during model training accordingly.
• Figure 2: The figure shows the obtained high-frequency image, along with the RGB images rendered by NeuS and Fre-NeuS, respectively. It can be seen that FreNeuS can improve the reconstruction accuracy of texture details by introducing high-frequency information, as demonstrated in features like the roof tile details in the image.
• Figure 3: Visualisation comparisons on different scenes of the DTU dataset. First column: reference images. Second to fifth columns: NeuS, HF-NeuS, 3DGS, 2DGS, and OURS. Chamfer distance errors are labelled in the lower left corner, with smaller metrics being better. Our method achieves minimal error while effectively capturing surface details. For instance, in the fourth row of the bear scene, Fre-NeuS successfully reproduces both the tooth gap and the distinct leg folds.
• Figure 4: Qualitative evaluation on the Ship and Material scenes. First column: reference images. Second to the fourth column: NeuS, HF-NeuS, and OURS. Compared to other methods, the inner circle of the sphere reconstructed by FreNeuS is more visible and the sail texture is much clearer.
• Figure 5: Visualisation of ablation experiments to verify the performance of each module. First column: the reference image. Second to fifth columns: NeuS, use of high-frequency constraints, use of high-frequency dynamic sampling strategies, and FreNeuS. It can be observed that as the two modules are added, the reconstructed surface details become clearer and more refined.