AIMS Seminar 21st January 2022
Retrieving 3D geometry from light reflection; from classical photometric stereo to deep learning based approaches
This work is related to Fotios’ best industrial paper at BMVC 2020 on DNN-based Photometric stereo.
Abstract - Reconstructing the 3D shape of an object using several images under varied illumination is a well-studied computer vision problem. Photometric Stereo (PS), exploits the relationship between shading and local shape and has traditionally been one of the most successful techniques at recovering a large amount of surface details.
However, up until recently, PS has been highly impractical and most approaches were only applicable in a very controlled lab setting and limited to objects experiencing diffuse reflection.
Moreover, complicated illumination effects such as point light propagation, cast shadows and self-reflections have been hard to deal with.
In this talk, recent advances in photometric stereo modelling will be reviewed. Latest findings have allowed to extend its applicability to most objects and surfaces under general illumination. These advances include differential modelling for variational optimisations as well as recent deep models. More reliable render engines have boosted the capability of providing synthetic data used by neural networks reliably. In particular, we have shown how to generate a vast amounts of synthetic training data that approximate the full set of relevant physical effects and thus allow training deep networks that achieve state of the art performance on all the available real test data.
BIO - Fotios Logothetis received the B.E. degree in 2014 and the Ph.D. degree in computer vision from the University of Cambridge in 2019 under the supervision of Prof Roberto Cipolla. He is currently working as a researcher at the Toshiba Cambridge Research Laboratory and his research interests include photometric stereo, shape-from-X and multi- view stereo.