Intrinsic Appearance Decomposition Using Point Cloud Representation

The aim of intrinsic decomposition is to deduce the albedo and shading components, typically from 2D images. However, this task is ill-posed, necessitating previous methods to rely on imaging assumptions. In contrast to 2D images, point clouds present a promising solution due to their richness as scene representation formats. They inherently align both the geometric and color information of an image, making them valuable to address this challenging problem. Hence, we propose a method, Point Intrinsic Net (PoInt-Net), which jointly predicts the albedo, light source direction, and shading by leveraging point cloud representations. Through experiments, we demonstrate the advantages of PoInt-Net, as it outperforms 2D representation methods across multiple metrics and datasets. Moreover, the model exhibits reasonable generalization capabilities for previously unseen objects and scenes.