基于重光照自监督约束的单张图像逆渲染方法
Single-Image Inverse Rendering with a Self-Supervised Relighting Constraint
刘兆瑞 1万丽莉1
作者信息
- 1. 北京交通大学计算机科学与技术学院,北京 100044;北京交通大学信息科学研究所,北京 100044
- 折叠
摘要
逆渲染是从图像中推断几何、光照和材质的过程。传统方法严重依赖多视角图像、特殊光照条件或精确的三维监督信息,限制了其在现实场景中的适用性。针对单张图像逆渲染面临的不适定性与标注数据匮乏问题,本文提出一种基于重光照自监督约束的逆渲染方法。该方法以自监督训练为核心,通过正向渲染模拟图像生成物理过程,训练网络学习逆向解耦逻辑,并引入重光照本征属性一致性约束,确保分解结果的物理合理性与视觉保真度。具体地,基于 Phong 光照模型构建可微分渲染链路,设计多分支网络实现几何、材质、光照参数的联合解耦；进一步构建重光照一致性框架,利用随机光照条件重新渲染解耦出的场景属性,约束两次分解的法线、反照率保持一致。在Co3D 真实数据集与COSy 合成数据集上的实验表明,本方法在法线和反照率解耦任务上均优于现有对比方法,且对未知类别及复杂光照场景具有强鲁棒性。
Abstract
Inverse rendering aims to infer geometry, illumination, and material properties from images. Traditional methods heavily rely on multi-view images, specialized lighting conditions, or precise 3D supervision, limiting their applicability to real-world scenarios. To address the ill-posed nature and scarcity of annotated data in single-image inverse rendering, we propose a relighting-based self-supervised inverse rendering method. Our approach centers on self-supervised training: we simulate the physical image formation process through forward rendering to train networks in learning inverse decomposition logic, and introduce a relighting intrinsic consistency constraint to ensure both physical plausibility and visual fidelity of the decomposed results. Specifically, we construct a differentiable rendering pipeline based on the Phong illumination model and design a multi-branch network for joint disentanglement of geometry, material, and lighting parameters. Furthermore, we establish a relighting consistency framework that re-renders the decomposed scene attributes under randomized lighting conditions, enforcing consistency in the recovered normals and albedos across both decomposition stages. Experiments on the real-world Co3D dataset and the synthetic COSy dataset demonstrate that our method outperforms existing approaches in normal and albedo disentanglement tasks, exhibiting strong robustness to unseen categories and complex lighting scenarios.关键词
逆渲染/重光照一致性/自监督学习Key words
Inverse Rendering/Relighting Consistency/Self-Supervised Learning引用本文复制引用
刘兆瑞,万丽莉.基于重光照自监督约束的单张图像逆渲染方法[EB/OL].(2026-04-27)[2026-04-28].http://www.paper.edu.cn/releasepaper/content/202604-201.学科分类
计算技术、计算机技术
评论