针对高亮物体的高效神经隐式表面重建

Efficient Neural Implicit Surface Reconstruction for Glossy Objects

摘要: 神经隐式表面(Neural Implicit Surface)一直是近年来计算机视觉的热门研究方向。许多工作通过扩展神经辐射场的体渲染管线实现了从二维图像和相机位姿作为输入，在无需三维监督信息下重建高质量三维物体形状。但是，由于利用二维图像进行训练监督，这些方法难以对高亮物体的形状进行合理的推理，其原因是因为物体材质和环境光照所产生的模糊性。本文提出一种针对高亮物体的高效表面重建方法，通过权重插值辐射场(Radiance Field)与反射场(Reflection Field)的方式使得可以更好地表达高亮物体的外观。同时，本文引入了一种渲染损失的方法来缓解高光反射带来的多视角不一致问题，并且引入了两种针对物体法向量的正则化来缓解混合神经场梯度噪声的问题。本工作通过渐进式的训练范式分别对三种数据集进行了实验，实验表明，本方法在多视角合成和高亮物体表面重建任务上都超越了基准模型，并且在训练速度上比基准模型快一个量级。

Abstract: Neural implicit surface has been a popular research direction in computer vision in recent years. Many approaches have extended the volume rendering pipeline of neural radiance field to reconstruct high quality 3D object shapes from 2D images and camera poses as input without any 3D supervision. However, due to the use of two-dimensional images for training supervision, it is difficult for these methods to rationally reason about the shape of glossy objects, because of the ambiguity caused by the material and environment lighting. In this paper, we propose an efficient surface reconstruction method specifically designed for glossy objects, which better represents the appearance of such objects through the interpolation of a weighted radiance field and reflection field. Additionally, we introduce a relax rendering loss to alleviate the issue of multi-view inconsistency caused by specular reflections, and two types of regularization for object normal to reduce the gradient noise of the hybrid neural field. Experiments on three datasets using a progressive training paradigm demonstrate that the proposed method outperforms baseline models in novel view synthesis and surface reconstruction tasks, while achieving training speeds approximately one order of magnitude faster than baseline models.