M-plicits：基于嵌套多尺度残差的神经隐式表面
原标题：M-plicits: Neural Implicit Surfaces via Nested Multiscale Residuals
AI 摘要
该论文提出 M-plicits，一种将表面建模为多层感知机（MLP）残差和的多尺度隐式神经表示框架，通过嵌套邻域训练策略将监督严格限制在前一层零水平集附近的窄带内。粗网络充当低通滤波器建立干净几何先验，后续残差逐步细化几何而不拟合高频噪声。在 Stanford 和 Thingi32 数据集上，M-plicits 在粗配置下取得最佳平均 Chamfer 距离，在细配置下取得最佳中位 Chamfer 距离和 IoU，噪声鲁棒性显著优于 iNGP、BACON 和 IDF，且参数量比基于网格的基线少一个数量级。作者还引入多尺度球体追踪和基于 GEMM 的解析法线计算，实现实时高保真渲染，代码、模型和数据将在 GitHub 发布。
正文节选
M-plicits: Neural Implicit Surfaces via Nested Multiscale Residuals 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