Génération automatique de cartes de profondeur relative par utilisation des occlusions dynamiques

Abstract

The lack of 3D content is a major obstacle to 3D television expansion. One viable solution to this obstacle is the 3D content generation out of ordinary 2D content. As a matter of fact, several depth cues can be found in a 2D image or video, which makes the automatic conversion from 2D to 3D possible. Among those cues, dynamic occlusions, that allow a relative order attribution to adjacent objects, represent a reliable base for 2D to 3D conversion and are available in all types of scenes. In this thesis, an improved 2D to 3D conversion approach is presented. It relies on the dynamic occlusion depth cue to generate relative depth maps. The front and back motion analysis between two consecutive frames allows the calculation of the dynamic occlusions. The utilized motion is calculated using a revised version of the optical flow estimation algorithm Epic-Flow, and the revisions added to this optical flow made it coherent in Forward-backward without compromising its performance. Thanks to this new feature, occlusions are simpler to calculate than the approaches used in the literature. Indeed, contrary to the approach followed by Salembier and Palou in 2014, the proposed method of calculating occlusions does not require the costly operation of motion estimation by region according to a quadratic model. The obtained occlusions allow order deduction of the objects contained in the image. These objects are obtained using segmentation, which considers both color and motion. The proposed method allows the automatic generation of relative depth maps in the presence of motion of the objects in the scene. It makes it possible to obtain results comparable to those obtained by Salembier and Palou, without requiring the estimation of movement by region.

Date	1 Mar 2018
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Carlos Vázquez (Supervisor) & Stéphane Coulombe (Co-supervisor)