Bonjour à tous,

Vous êtes invité à assister à l'examen Prédoc III de Chaima El Asmi, jeudi le 19 décembre, à 14h.

Jury

Abstract

Motion estimation, is a key task in computer vision, aimed at estimating motion information from video sequences. Optical flow, a sub-problem of motion estimation, estimates a dense motion field, one vector for each pixel. This work explores the learning process with regard to features for correspondence matching, in the context of realistic scenes, various levels of flow parametrization, and small vs.large displacements. We propose to use the feature grid introduced for neural radiance field to structure cost functions and start bridging the gap between continuous and discrete methods. In that regard, we successfully tested a continuous architecture integrating the traditional Lucas & Kanade approach.

Initial experiments demonstrate that feature grids can structure a latent feature representation for normal flow regression as well as for optical flow correspondence. The aim at hand is to connect continuous and discrete matching in a coherent and compatible framework through the use of structured continuous feature representations, optimal scale and proper flow prediction. Overall, we hope to achieve a universal motion flow architecture that will be adaptive to image scale and apply to a wide range of image sequences.