International organisation for standardisation organisation internationale de normalisation

5.1.53D-AVC-CE5: Depth range compensation for inter/inter-view prediction

m23604

Teruhiko Suzuki

3D-AVC-CE5 summary report

Summary report: The CE was planned to investigate depth range compensation for both inter and inter-view prediction, but tools are already supported in test model under consideration. It was concluded that further analysis should be done as part of an AHG. There also seemed to be a relationship to high-level syntax design.

Conclusion:

• 
  Discontinue CE, but continue this work as part of AHG study
  

5.1.63D-AVC-CE6: In-loop depth resampling

m23768	Thomas Rusert	3D-AVC-CE6 summary report
m23661	Seok Lee, Seungsin Lee, HoCheon Wey, Jaejoon Lee	3D-AVC-CE6 related results on Samsung’s in-loop depth resampling
m23673	Seok lee, Jaejoon Lee, Du Sik Park	Cross check of 3D-AVC-CE6 Ericsson’s adaptive depth map upsampling by Samsung
m23721	Payman Aflaki Beni, Dmytro Rusanovskyy, Miska M. Hannuksela	Non-linear Depth Map Resampling for 3DV-ATM Coding
m23766	Ivana Girdzijauskas, Kenneth Andersson, Thomas Rusert	3D-AVC-CE6 results on edge-adaptive median filtering
m23767	Ivana Girdzijauskas, Thomas Rusert	Cross-check of Mitsubishi’s 3D-AVC-CE6 results
m23774	Danillo Graziosi, Dong Tian, Anthony Vetro	3D-AVC-CE06 results on in-loop depth resampling by Mitsubishi
m23775	Dong Tian, Danillo Graziosi, Anthony Vetro	Cross-check 3D-AVC-CE06 report on Ericsson’s proposal for in-loop depth resampling

Summary report:

• 
  The CE investigated depth resampling in the case that depth samples are used for normative aspects such as view synthesis prediction.
  
• 
  Ericsson used down-sampling as in the reference, but modified the depth up-sampling using bilinear interpolation. This proposal also uses texture edge detection.
  
• 
  Mitsubishi proposed to modify the down-sampling based on median filtering and to perform up-sampling based on a frequent-low-high filter and median filter. This had a large increase in decoding complexity.
  
• 
  The CE report recommended subjective viewing to verify the results. It also suggested that further CE work should study different factors including down-sampling, in-loop up-sampling and post-processing.
  

m23661 (Samsung) was a CE6 related contribution that proposed 3x3 gray scale dilation filtering after linear interpolation to reduce synthesis artifacts in up-sampled depth maps. Reported bit rate reductions were 7.2% and 7.7% in synthesized views for 2-view and 3-view cases, with a complexity increase of 120-126%. A maximum of 22% gain was reported for "Undo Dancer", with good visual improvement. Further evaluation of this in CE6 was recommended.

m23721 (Nokia) was a CE6 related contribution that proposed new down-sampling and up-sampling operations. A 7.4% bit rate reduction was reported on on average, and up to 20% for "Undo Dancer" with good visual improvement. Up-sampling alone reportedly yields approximately 3.3% gain, with 50% higher complexity. This was proposed as a pre/post-process and it was requested that CE work should clarify the gains of in-loop filtering and the other filtering that is done outside the loop.

m23810 (Qualcomm) was a CE6 related contribution that proposed adaptive depth edge filtering, where only edge-like pixels in edge-like MBs are filtered. It also applied smoothing along depth edges to reduce ringing artifacts. Approximately 7% gain was reported on synthesized views, with up to 22% on "Undo Dancer" with notable visual improvement. The associated encoder complexity increase is 162-139%, and the decoder complexity increase is 105-107%.

Subjective viewing was conducted during the week. No major differences were observed.

Conclusion:

• 
  Continue CE.