International organisation for standardisation organisation internationale de normalisation

5.1.103D-AVC-CE10: Depth In-loop Filtering

m23684	Jaejoon Lee	3D-AVC-CE10 summary report on depth in-loop filtering
m23664	Ilsoon Lim, Jaejoon Lee, Du Sik Park	3D-AVC-CE10 results on Samsung's region-based adaptive loop filter
m23675	Ilsoon Lim, Jaejoon Lee, Du Sik Park	Cross check of 3D-AVC-CE10 related Qualcomm's depth in-loop filter by Samsung
m23810	Rong Zhang, Ying Chen, Marta Karczewicz	3D-AVC-CE10 Results on Depth In-loop Filtering by Qualcomm
m23859	Rong Zhang, Ying Chen	Cross check of 3D-AVC-CE10 Results on Depth In-loop Filtering by Samsung

Summary report:

Samsung proposed a region based adaptive bilateral filter, with reported 0.9% and 1.25% gains without JVF and when combined with JVF, for 2-view case, and 0.74% and 1.67% gains for 3-view. Complexity increases of approximately 30% in the encoder and 18~40% in the decoder were reported – mainly due to determining the filter coefficients for each slice.

Qualcomm proposed an adaptive depth edge filter and smoothing along depth edges. This was considered to be outside of the scope of this CE and related to CE6.

Proponents recommended adopting region-based adaptive bilateral filter.

The group discussed gain vs. complexity, and considered the potential to reduce complexity.
A BOG discussion was held on Low Complexity In-Loop Operations, and a BOG report was produced (m23992). Its recommendations included:

1. 
   Adopt Qualcomm-Zhejiang reduced complexity VSP after the cross-check (Nokia to cross-check).
   
   1. 
      Remove boundary detection
      
   2. 
      Remove noise cleaning process
      
   3. 
      Single-loop warping and hole filing
      
2. 
   Zhejiang University is expected to take Qualcomm software in use and introduce "merge of splatting & warping".
   
   1. 
      Qualcomm shares the software with Zhejiang University and Nokia
      
   2. 
      Zhejiang University software is to be delivered by 17 Feb. 2012 to Nokia
      
   3. 
      Nokia will cross-check the following:
      
      1. 
         Encoder/Decoder match
         
      2. 
         Decoder complexity reduction vs. anchor: >15% of execution time, with penalty < 0.4% of BDR for decoded views
         
      3. 
         If cross-check confirm this results, software will become 3DV-TM v0.3
         
      4. 
         If cross-check 17.Feb.2012 failed, but cross-check from Qualcomm is OK, Qualcomm software will become 3DV-TM v0.3
         
3. 
   Zhejiang University and others are encouraged to Study in CE1 sub-pixel accuracy of the warping process in VSP.
   
4. 
   The CE on in-loop resampling will continue, and proponents are encouraged to revisit the design, and take complexity in the account.
   
5. 
   The CE on depth inloop filtering will continue, and proponents are encouraged to revisit the design, and take complexity in the account.
   
6. 
   Studying the decoding time complexity estimate with disabled writing of output files to the disk.
   
7. 
   Studying the checksum output of the decoder, which would enable disabling decoder output.
   
8. 
   The mode of Low Resolution DMVP is not necessary at this stage.
   

Conclusions:

• 
  Continue CE
  
• 
  BoG recommendations agreed as described above.