Learning advection and transfer to another domain
Encadrants : Dominique Béréziat
Email : dominique.bereziat@lip6.fr
Disponible : OUI
Spécialité : IMA/DIGIT
Nombre d'étudiants : 1
Description : Estimating apparent motion (also named optical flow) from 2 or more images is one of the standard computer field problems.
Nowadays, the best methods are based on DL models but need ground truth to be trained. Learning optical flow in an unsupervised remains possible but with weak performance. In this project, we propose to learn the advection process underlying optical flow in a supervised framework, using the large datasets available in the literature. Having learned advection, we propose in a second step to train an optical flow estimator in an unsupervised framework. In a final step, we aimed to transfer to an another domain, such as ocean surface circulation
Pré-requis : pytorch
Travail demandé : - review of the literature: old-school optical flow estimation, deep learning optical flow estimation,
- advection learning: implementation, training, and testing with a basic model (Unet ?)
- transfer: use of Flownet combined with learned advection, train, test and generalisation
