Information Theoretic Modeling of High Precision Disparity Data for Lossy Compression and Object Segmentation

Abstract

In this paper, we study the geometry data associated with disparity map or depth map images in order to extract easy to compress polynomial surface models at different bitrates, proposing an efficient mining strategy for geometry information. The segmentation, or partition of the image pixels, is viewed as a model structure selection problem, where the decisions are based on the implementable codelength of the model, akin to minimum description length for lossy representations. The intended usage of the extracted disparity map is to provide to the decoder the geometry information at a very small fraction from what is required for a lossless compressed version, and secondly, to convey to the decoder a segmentation describing the contours of the objects from the scene. We propose first an algorithm for constructing a hierarchical segmentation based on the persistency of the contours of regions in an iterative re-estimation algorithm. Then, we propose a second algorithm for constructing a new sequence of segmentations, by selecting the order in which the persistent contours are included in the model, driven by decisions based on the descriptive codelength. We consider real disparity datasets which have the geometry information at a high precision, in floating point format, but for which encoding of the raw information, in about 32 bits per pixels, is too expensive, and we then demonstrate good approximations preserving the object structure of the scene, achieved for rates below 0.2 bits per pixels.