FAIR3D: Frugal and explainable Artificial intelligence for photometric 3D-Reconstruction

Keywords: computer vision, 3D-reconstruction, deep learning

Supervisors: David Tschumperlé and Yvain Quéau, CNRS research fellows

Location: Image team, GREYC laboratory, Caen, France

Net salary: approx. 1700€/month

Starting date: autumn 2024

Application: https://doctorat.campusfrance.org/CF202436799

Scientific program:

The proposed project focuses on addressing key challenges in the field of photometric 3D reconstruction, also known as photometric stereo, which entails estimating surface relief from photographs acquired under active lighting conditions. Traditionally, this problem has been approached through the numerical inversion of a parametric physical model, yielding cost-effective and interpretable solutions adept at reconstructing fine surface structures approximated well by the Lambertian reflectance model for matte surfaces. However, recent years have seen a shift towards artificial intelligence-based approaches, where neural network architectures directly learn the association between images and geometry from massive datasets. While such methods excel in reconstructing relief for highly reflective surfaces, they pose challenges in terms of resource requirements and interpretability, as reflectance models are implicitly encoded within numerous opaque network coefficients.

This project aims to tackle these challenges by developing artificial intelligence methods that are both resource-efficient and interpretable for photometric 3D reconstruction. The primary objective is to design a neural reflectance model using lightweight networks and to embed this architecture within a classical inverse problem approach. This integration will yield a state-of-the-art algorithmic solution that effectively balances performance, computational efficiency and interpretability of results.

Required background and skills

We are looking for a MSc student with a strong background in applied computer science, in particular imaging. Good skills in machine learning and optimization, and a basic knowledge of computer vision, would be a plus. The candidate should also show good programming skills, and good communication skills in English, both written and oral.

Application

Applications through the CampusFrance portal : https://doctorat.campusfrance.org/CF202436799