Multiscale neural models improve 3D surface reconstruction speed and quality
M-plicits: Neural Implicit Surfaces via Nested Multiscale Residuals
Computer Vision and Pattern RecognitionGraphicsMachine Learning
Summary
Representing 3D shapes using neural networks can be slow and sensitive to noisy data. The authors propose M-plicits, which uses multiple neural networks working together at different scales, each focusing on refining details only near the current surface. This approach keeps the shape smooth and resistant to noise while making rendering faster using a new method. Their tests show this method is both accurate and efficient compared to older techniques.
What this means in practice
- •For 3d graphics developers: Render detailed 3D surfaces in real-time with fewer parameters and less noise sensitivity for applications like games or simulations.
- •For computer-aided design teams: Improve accuracy and speed when reconstructing noisy or incomplete 3D scans for design and manufacturing workflows.
Authors
Vinícius da Silva, Isabelle Melo, Matheus Bessa, Guilherme Schardong, Luiz Schirmer, André Araújo, Nuno Gonçalves, Hélio Lopes, Alberto Raposo, Luiz Velho, Tiago Novello
Abstract
Encoding input coordinates with sinusoidal functions into multi-layer perceptrons (MLPs) has proven effective for implicit neural representations (INRs) of surfaces defined as zero-level sets. However, existing methods often struggle to balance training efficiency, rendering speed, and noise robustness: single-MLP approaches are expensive at inference, grid-based representations are fast but can limit surface smoothness and overfit input noise, and previous multiscale approaches frequently capture noise and produce artifacts due to hard spectral truncation. To address these limitations, we propose M-plicits, a multiscale framework that models surfaces as a residual sum of MLPs trained via a sequence of nested neighborhoods. Unlike existing residual approaches that rely on standard domain-wide sampling and require costly mesh extraction for visualization, our method strictly localizes supervision to narrow bands around the previous zero-level sets. This nested design naturally provides robustness against noisy input data: the coarse network acts as a low-pass filter that establishes a clean geometric prior, while subsequent residuals progressively refine the geometry without fitting to high-frequency artifacts. We further introduce a multiscale sphere-tracing algorithm and a GEMM-based analytical normal computation that bypasses auto-differentiation entirely, yielding high-fidelity real-time rendering. On Stanford and Thingi32, M-plicits achieves the best mean Chamfer distance in the coarse configuration and the best median Chamfer distance and IoU in the fine configuration, with substantially better noise robustness than iNGP, BACON, and IDF, while using an order of magnitude fewer parameters than grid-based baselines. Code, models, and data will be released at https://github.com/dsilvavinicius/m-plicits.