Foveated 3D range geometry compression via loss-tolerant variable precision depth encoding

Abstract

The capacity of three-dimensional (3D) range geometry acquisition methods to capture high-precision scans at high frame rates increases every year. These improvements have influenced a broadening range of disciplines to implement 3D range geometry capture systems, including telepresence, medicine, the visual arts, and many others. However, its increased popularity, precision, and capture rates have caused mounting pressure on the storage and transmission of 3D range geometry, thus straining their capacities. Compression techniques seek to alleviate this pressure by offering reduced file sizes, while maintaining the levels of precision needed for particular applications. Several such compression methods use sinusoidal modulation approaches to encode floating-point 3D data into conventional 2D red, green, and blue (RGB) images. In some applications, such as telepresence, high precision may only be required in a particular region within a depth scan, thus allowing less important data to be compressed more aggressively. This paper proposes a feature-driven compression method that provides a way to encode regions of interest at higher levels of precision while encoding the remaining data less precisely to reduce file sizes. This method supports both lossless and lossy compression, enabling even greater file-size savings. For example, in the case of a depth scan of a bust, an algorithmically extracted bounding box of the face was used to create a foveated encoding distribution so that the facial region was encoded at higher precisions. When using JPEG 80, the RMS reconstruction error of this novel, to the best of our knowledge, encoding was 0.56 mm in the region of interest, compared to a globally fixed higher precision encoding where the error was 0.54 mm in the same region. However, the proposed encoding achieved a 26% reduction in overall compressed file size compared to the fixed, higher-precision encoding.