Anti-Aliased Neural Implicit Surfaces with Encoding Level of Detail

Abstract: We present LoD-NeuS, an efficient neural representation for high-frequency geometry detail recovery and anti-aliased novel view rendering. Drawing inspiration from voxel-based representations with the level of detail (LoD), we introduce a multi-scale tri-plane-based scene representation that is capable of capturing the LoD of the signed distance function (SDF) and the space radiance. Our representation aggregates space features from a multi-convolved featurization within a conical frustum along a ray and optimizes the LoD feature volume through differentiable rendering. Additionally, we propose an error-guided sampling strategy to guide the growth of the SDF during the optimization. Both qualitative and quantitative evaluations demonstrate that our method achieves superior surface reconstruction and photorealistic view synthesis compared to state-of-the-art approaches.

• The paper presents LoD-NeuS, a novel method integrating multi-scale tri-plane encoding and cone tracing for anti-aliased 3D surface rendering.
• It employs error-guided SDF growth for precise reconstruction of thin structures and intricate geometric details.
• Quantitative results, including improvements in PSNR and Chamfer Distance, demonstrate its superiority over state-of-the-art techniques.

Anti-Aliased Neural Implicit Surfaces with Encoding Level of Detail

Introduction

The paper introduces LoD-NeuS, a neural representation focusing on high-frequency geometry detail recovery and anti-aliased novel view rendering. By leveraging the concept of Level of Detail (LoD), commonly utilized in voxel-based models, it proposes an innovative approach to enhance the quality of 3D reconstructions and visual renderings. Through a multi-scale tri-plane-based representation, LoD-NeuS aggregates space features efficiently and applies differentiable rendering to optimize the LoD feature volume. Additionally, it adapts an error-guided sampling technique to facilitate the growth of the signed distance function (SDF), ensuring improved surface reconstruction.

Multi-Scale Tri-plane Encoding

LoD-NeuS builds upon the tri-plane representation, which consists of three orthogonal feature planes. It introduces multi-scale tri-planes to capture varying levels of geometric detail. This involves projecting a 3D point onto these planes and interpolating features from them. Unlike previous methods that tend to use fixed frequency position encoding, LoD-NeuS provides a concatenated feature vector from multiple levels of tri-planes, offering a robust and adaptable encoding mechanism that enhances surface detail reconstruction.

Anti-Aliasing and Volume Rendering

Traditional renderings typically miss accounting for pixel size when casting rays, leading to aliasing issues. LoD-NeuS tackles this by reformulating volume rendering via cone tracing, which considers the pixel as a cone with conical frustums. Through Cone Discrete Sampling, it effectively integrates geometric and color information, using neighboring vertices within the conical frustum. Furthermore, Multi-convolved Featurization approximates Gaussian integration within these frustums, offering a continuous and efficient representation for high-frequency details.

Figure 1: Aggregation of LoD feature, including multi-convolved featurization and cone discrete sampling.

SDF Growth and Refinement

The paper observes difficulties in reconstructing thin surfaces with SDF due to rapid changes in signed distances. To mitigate this, an SDF Growth Refinement strategy is proposed, where initial error maps guide the sampling process. Starting from detected growth points, the method incrementally expands regions to refine the SDF, enhancing the accuracy and completeness of thin structures.

Figure 2: Comparison of SDF growth method against random selection of rays around the error map.

Results and Comparisons

LoD-NeuS displays superior performance in achieving high fidelity geometry and photorealistic novel view synthesis. Quantitative results, such as PSNR and Chamfer Distance, highlight its improvements over state-of-the-art approaches like NeuS, HF-NeuS, and NeRF. It successfully reproduces intricate details and minimizes aliasing artifacts while maintaining geometric integrity, visible in detailed mesh reconstructions and zoom-in comparisons.

Figure 3: Qualitative comparison with zoom-in details showing superior performance of LoD-NeuS in novel view synthesis.

Figure 4: Comparison of novel-view synthesis and mesh reproduction with details surpassing HF-NeuS and NeuS.

Conclusion

LoD-NeuS presents a noteworthy advancement in neural rendering by integrating continuous levels of detail encoding with efficient featurization and rendering methods. It proves its capability in reconstructing complex geometries while avoiding aliasing, all within a feasible computational framework. Future work may focus on further optimizing the SDF growth refinement and exploring applications beyond the current datasets, providing broader adaptability and insight into continuous LoD neural surfaces.