Open Call for european P-SPHERE project researcher (Post Doc): Predicting intrinsic properties from images using deep networks<![CDATA[TOPIC DESCRIPTION Natural images contain many effects caused by the interaction between light and the objects in the scene. These effects include, amongst others, shading, shadows, and pecularities. In general, computer vision algorithms are hindered by such effects. Identifying and isolating these effects and also other physical elements of the scene which participate in the generation of these effects, such as 3D shape, depth and the color of the light source, can be very useful to improve the performance of many algorithms. However, the estimation of these intrinsic properties from a single image is a challenging problem which has received much attention in the last years. This project aims at the definition of deep network architectures that allows to decompose images into some of their photometric properties such as reflectance, lighting, shading (including self-shadows and cast shadows) or peculiarities jointly with the corresponding 3D shape properties The fellow is expected to contribute in the current lines of research of the hosting group, and do research tasks as:
- Reviewing previous works on the topics of the project (i.e. intrinsic image estimation and deep learning).
- Designing Convolutional Neural Networks to learn the best features based on the large dataset of lighting effects that the hosting group is building.
- Evaluating the defined models by performing experiments on standard datasets and/or specific datasets of intrinsic properties.