Visual and Geometric Data Compression for Immersive Technologies

Abstract

The contributions of this thesis are new compression algorithms for light field images and point cloud geometry. Light field imaging attracted wide attention in the recent decade, partly due to emergence of relatively low-cost handheld light field cameras designed for commercial purposes whereas point clouds are used more and more frequently in immersive technologies, replacing other forms of 3D representation. We obtain successful coding performance by combining conventional image processing methods, entropy coding, learning-based disparity estimation and optimization of neural networks for context probability modeling.

On the light field coding side, we develop a lossless light field coding method which uses learning-based disparity estimations to predict any view in a light field from a set of reference views. On the point cloud geometry compression side, we develop four different algorithms. The first two of these algorithms follow the so-called bounding volumes approach which initially represents a part of the point cloud in two depth maps where the remaining points of the cloud are contained in a bounding volume which can be derived using only the two depth maps that are losslessly transmitted. One of the two algorithms is a lossy coder that reconstructs some of the remaining points in several steps which involve conventional image processing and image coding techniques. The other one is a lossless coder which applies a novel context arithmetic coding approach involving gradual expansion of the reconstructed point cloud into neighboring voxels. The last two of the proposed point cloud compression algorithms use neural networks for context probability modeling for coding the octree representation of point clouds using arithmetic coding. One of these two algorithms is a learning-based intra-frame coder which requires an initial training stage on a set of training point clouds. The lastly presented algorithm is an inter-frame (sequence) encoder which incorporates the neural network training into the encoding stage, thus for each sequence of point clouds, a specific neural network model is optimized which is also transmitted as a header in the bitstream.