Surface Normal Reconstruction From Specular Information in Light Field Data

Marcel Gutsche, Hendrik Schilling, Maximilian Diebold, Christoph Garbe; Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017, pp. 22-29

Abstract


Specular highlights provide information about the shape of an object. Its characteristics are mostly unwanted in computer vision due to violation of the Lambertian assumption, which most algorithms require. Instead of neglecting this ubiquitous phenomenon we harvest it to extract surface normals with very high accuracy. Compared to photometric stereo our method works with multiple views and a fixed light source. We only require a low number of observation from a small part of the specular lobe to reconstruct the normal and reflection parameters. This is achieved by jointly optimizing the normal and the light transport of surfaces points. This work is a proof of concept to demonstrate the feasibility of acquiring highly accurate surface normals from specular reflection, which can be combined with conventional methods. The model is tested and evaluated on synthetic as well as real world data acquired by a cross light field setup.

Related Material


[pdf]
[bibtex]
@InProceedings{Gutsche_2017_CVPR_Workshops,
author = {Gutsche, Marcel and Schilling, Hendrik and Diebold, Maximilian and Garbe, Christoph},
title = {Surface Normal Reconstruction From Specular Information in Light Field Data},
booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {July},
year = {2017}
}