Volumetric depth confidence's denoising.

Confidence assignment to the depth voxel.

In this paper, we define volumetric depth confidence and propose a method to denoise this data by performing adaptive wavelet thresholding using three dimensional (3D) wavelet transforms. The depth information is relevant for emerging interactive multimedia applications such as 3D TV and free-viewpoint television (FTV). These emerging applications require high quality virtual view rendering to enable viewers to move freely in a dynamic real world scene. Depth information of a real world scene from different viewpoints is used to render an arbitrary number of novel views. Usually, depth estimates of 3D object points from different viewpoints are inconsistent. This inconsistency of depth estimates affects the quality of view rendering negatively. Based on the superposition principle, we define a volumetric depth confidence description of the underlying geometry of natural 3D scenes by using these inconsistent depth estimates from different viewpoints. Our method denoises this noisy volumetric description, and with this, we enhance the quality of view rendering by up to 0.45 dB when compared to rendering with conventional MPEG depth maps.

Publications

• Srinivas Parthasarathy, Akul Chopra, Emilie Baudin, Pravin Kumar Rana, and Markus Flierl:
  Denoising of Volumetric Depth Confidence for View Rendering,
  In Proceedings of 3DTV-Conference The True Vision: Capture, Transmission and Display of 3D Video (3DTV CON), Zurich, Switzerland, October 2012.