The paper proposes Spin-UP, an unsupervised method to address the natural light uncalibrated photometric stereo (NaUPS) problem. NaUPS aims to reconstruct surface normals given images of an object captured under arbitrary environment light, relieving the strict assumptions in classical uncalibrated photometric stereo methods.
The key highlights of the paper are:
Novel image capture setup: Spin-UP uses a rotatable platform to capture images of the object, which reduces the unknowns in the light representation and provides reliable priors to alleviate the ill-posedness and ambiguities in NaUPS.
Light initialization method: Leveraging the object's occluding boundaries, Spin-UP derives a reliable initial environment light model to mitigate the ambiguity between light and object properties.
Advanced light and material models: Spin-UP adopts a spherical Gaussian model for the environment light and a modified Disney BRDF model for the spatially varying and isotropic reflectance, enabling it to handle a broader range of scenarios compared to previous methods.
Unsupervised optimization: Spin-UP recovers the surface normals, environment light, and isotropic reflectance through an iterative unsupervised optimization process based on neural inverse rendering.
Training strategies: Spin-UP employs interval sampling and shrinking range computing to reduce the computational cost and improve convergence during training.
Experiments on synthetic and real-world datasets demonstrate that Spin-UP outperforms previous supervised and unsupervised NaUPS methods and achieves state-of-the-art performance in handling general objects under complex natural illumination.
A otro idioma
del contenido fuente
arxiv.org
Ideas clave extraídas de
by Zongrui Li,Z... a las arxiv.org 04-03-2024
https://arxiv.org/pdf/2404.01612.pdfConsultas más profundas