Interactive spatio-temporal exploration of massive time-Varying rectilinear scalar volumes based on a variable bit-rate sparse representation over learned dictionaries

Abstract

We introduce a novel approach for supporting fully interactive non-linear spatio-temporal exploration of massive time-varying rectilinear scalar volumes on commodity platforms. To do this, we decompose each frame into an octree of overlapping bricks. Each brick is further subdivided into smaller non-overlapping blocks compactly approximated by quantized variable-length sparse linear combinations of prototype blocks stored in a learned data-dependent dictionary. An efficient tolerance-driven learning and approximation process, capable of computing the tolerance required to achieve a given frame size, exploits coresets and an incremental dictionary refinement strategy to cope with datasets made of thousands of multi-gigavoxel frames. The compressed representation of each frame is stored in a GPU-friendly format that supports direct adaptive streaming to the GPU with spatial and temporal random access, view-frustum and transfer-function culling, and transient and local decompression interleaved with ray-casting. Our variable-rate codec provides high-quality approximations at very low bit-rates, while offering real-time decoding performance. Thus, the bandwidth provided by current commodity PCs proves sufficient to fully stream and render a working set of one gigavoxel per frame without relying on partial updates, thus avoiding any unwanted dynamic effects introduced by current incremental loading approaches. The quality and performance of our approach is demonstrated on massive time-varying datasets at the terascale.