Skip to Main Content
In many areas of science and engineering, the desires to study a problem at the highest possible resolution have led to an explosive growth of data. It is imperative to reduce the data to a manageable scale for analysis and visualization. For high-precision floating-point data, compressing the data solely based on values can only achieve a limited saving. Further reduction is possible with the fact that usually only a smaller subset of the data is of interest in analysis. In this paper, we present an application-driven approach to compressing large-scale time-varying volume data. Our method identifies a reference feature to partition the data into space-time blocks, which are compressed with various precisions depending on their association to the feature. Runtime decompression is performed with bit-wise texture packing and deferred filtering. We show that our method achieves high compression rates and interactive rendering while preserving fine details surrounding regions of interest. Such an application-driven approach points us to a promising direction for coping with the large data problems facing computational scientists.