Keyframe-based Depth Compression (Update)

Contact person: Dmitriy S. Vatolin (dmitriy@graphics.cs.msu.ru)

Proposed 3D video compression scheme

picture

Depth maps contain uniform areas and very good compression rates are achieved. Depth maps have very good correlation with 2D image so this information is used for 3D video compression.

Depth map compression: objective evaluation

  • Depth map compression
    • MSU 3D compression
    • x264
  • RGB video was uncompressed

picture

  • Improved key frames encoding (comparison of JPEG2000 implementation from FFmpeg and ImageMagick packages)
  • Reference codec for depth map: x264 r2183

picture
Source stereo sequence is provided by Alan Bovik’s group

Adaptive key frames selection

Reduce bitrate up to 50% at the constant quality

  • Initial version: constant intervals between key frames
  • Improvement: put additional key frames in the places with low metric

picture

picture

Stereo compression: visual evaluation

  • Unable to use per-pixel metrics
    Reasons:
    • Depth map estimation
    • Stereo generation
  • Reference codec: H.264 MVC from Intel Media SDK version 3.5.915.45249 (high quality preset)
  • Our approach
    • 1 view compression with reference codec
    • depth map compression with YUVsoft Depth Upscale, YUVsoft Depth Propagation and JPEG2000
  • The proposed method shows better visual results of compression especially for small objects and detailed textures
  • The highest quality gain was achieved on smaller bitrates (2 Mbps)

picture

picture
Left view from “Step Up” stereo sequence

picture
Left view from “Avatar” stereo sequence

Problems and further work:

  • Depth Propagation algorithm is of high computational complexity. Using propagation of downscaled key frames and Depth Upscale can significantly improve speed of processing
  • Construction of multiview video from depth maps. The final quality highly depends both from depth map generation way and from multiview from 2D+depth generation algorithm
  • Depth key frames are set at constant intervals now. Selection of position of keyframes according to features of the video should Improve quality and compression rate
  • Quality measurement.Traditional metrics like PSNR and SSIM do not cover all aspects of 3D Video quality

Projects

Keyframe-based Depth Compression

Keyframe-based Depth Compression with Depth Propagation for Multiview Video Coding

Team

    Acknowledgements

    This work is partially supported by the Intel/Cisco Video-Aware Wireless Network (VAWN) Program