Whitepaper: Analysis and comparison of the most common depth video compression techniques
Since the release of the first Kinect camera for Microsoft’s Xbox in 2010, researchers and industry alike have made use of the additional dimension of 3D video to simplify, improve and extend vision solutions in a wide range of industries.
Advances in compute power and the availability of public clouds, have enabled ever more powerful machine vision solutions.
However, challenges posed by 3D video such as
- novel data formats defying existing video compression,
- non-unified 3D camera interfaces and
- ever-increasing resolution and framerate are limiting the scalability of machine vision for 3D video.
This report analyses three widely used lossless (Lz4, PNG and RVL) and two relevant lossy compression schemes for depth data. Then we present Aivero’s proprietary depth compression solution, 3DQ.
This document first provides background into 3D data, then we describe the main challenges of 3D video compression and the most commonly used approaches before presenting our solution – 3DQ. Finally, we study the performance of these solutions across open datasets and present a short discussion about the results.
Our results show that 3DQ offers a flexible depth compression solution that can be catered to specific applications and setup, being able to achieve very high compression ratios while maintaining image quality.
Finally, Aivero’s 3DQ based RGBD Toolkit for 3D video capturing, streaming, and storage is presented.