Maceira M. Multi-view depth coding based on a region representation combining color and depth information. Ruiz-Hidalgo J, Morros JR. Signal Theory and Communications (TSC). Universitat Polit├Ęcnica de Catalunya (UPC); 2017.

Abstract

Depth map data is used to supplement the color data in multi-view sequences. As depth maps present distinct characteristics than natural color images, new coding techniques are required to represent their smooth regions and sharp edges. In this thesis, segmentation-based coding techniques are proposed to encode depth maps by exploiting the redundancy between color and depth information. Methods developed combine partitions obtained from color and depth images to find efficient representations. The color image is assumed to be available before the depth map coding process, therefore a color partition can be obtained at the decoder without introducing coding cost.

Two hierarchical image segmentation algorithms are proposed to generate color and depth partitions for coding applications. The color segmentation obtains a super-pixel representation using color information, spatial distribution and shape complexity. The depth segmentation uses a 3D planar model for each region to extract the structure of the scene. Color and depth partitions are combined in depth map coding methods to find the final coding partition.

Different methods for texture representation have been explored in this thesis. Initial approaches used 2D coding methods, while a 3D representation have been proposed to represent depth maps from multiple views with a unique segmentation. This 3D representation is used to segment depth maps in single-view and multi-view configurations. Final coding partitions are obtained with a ratedistortion optimization over a hierarchy of regions. Segmentation-based coding techniques proposed obtain competitive results with HEVC coding standards.