Section: Overall Objectives

Representation and compression of visual data

The objective is to develop algorithmic tools for constructing low-dimensional representations of multi-view video plus depth data, of 2D image and video data, of visual features and of their descriptors. Our approach goes from the design of specific algorithmic tools to the development of complete compression algorithms. The algorithmic problems that we address include data dimensionality reduction, the design of compact representations for multi-view plus depth video content which allow high quality 3D rendering, the design of sparse representation methods and of dictionary learning techniques. The sparsity of the representation indeed depends on how well the dictionary is adapted to the data at hand. The problem of dictionary learning for data-adaptive representations, that goes beyond the concatenation of a few traditional bases, has thus become a key issue which we address for further progress in the area.

Developing complete compression algorithms necessarily requires tackling visual processing topics beyond the issues of sparse data representation and dimensionality reduction. For example, problems of scalable, perceptual, and metadata-aided coding of 2D and 3D visual data, as well as of near lossless compression of medical image modalities (CT, MRI, virtual microscopy imaging) are tackled. Finally, methods for constructing rate-efficient feature digests allowing processing in lower-dimensional spaces, e.g. under stringent bandwidth constraints, also falls within the scope of this research axis.