International organisation for standardisation organisation internationale de normalisation

18Non-TE Technical Contributions

18.1Loop filtering

JCTVC-C085 Comparison of loop & post filtering for in-loop and post-processing filtering AHG [T. Yamakage, T. Chujoh, T. Watanabe (Toshiba)]

This contribution showed experimental results of comparison between in-loop filtering and post filtering. This was a response to one of the mandates of the In-loop and post-processing filtering AHG. Coding efficiency of several filtering schemes was compared for in-loop or post filtering and discussed.

Remarks recorded included the following:

• 
  The current QC_ALF of the TMuC performs 4% better than a similar stand-alone post-filter
  
• 
  The method from AVC (SEI) AAPF performs 2.5% worse than QC_ALF (an amount which slightly decreases when operated in the loop)
  
• 
  Results also reportedly show that a filter must be explicitly designed as a loop filter or post filter
  

JCTVC-C113 Adaptive loop filter with low encoder complexity [I.S. Chong, W. Chien, M. Karczewicz (Qualcomm)]

In this contribution, low complexity versions of the Adaptive Loop Filter (ALF) with at most two encoding passes were proposed. Simulation results for one-pass and two-pass ALF are presented to demonstrate that one-pass and two-pass methods significantly lower the complexity of the original ALF with small performance degradation (less than 0.1% for two-pass and less than 0.7% for one-pass).

• 
  The one-pass method derives the map by using a filter from the previous frame and then optimizing coefficients for the current frame based on regions where the filter is switched on.
  
• 
  A two-pass method with one more iteration was described
  
• 
  A post filter from the VCEG KTA software was also compared (losing 0.9% BR relative to the anchor)
  
• 
  The method is very similar to the one proposed in JCTVC-C082
  

JCTVC-C195 An enhanced block-based adaptive loop filter in TMuC0.7.0 platform [P. Wu, S. Paschalakis, N. Sprljan (Mitsubishi Electric)]

In this contribution an Enhanced Block-based Adaptive Loop Filter (EBALF) algorithm has been implemented and integrated into TMuC0.7.0 platform as a complete functional block for in-loop filtering processing. Without altering the current default adaptive loop filter technology, which is contributed by Qualcomm as QC_ALF in the TMuC software, this EBALF functional block can be placed before or after QC_ALF, resulting in an in-loop filtering composed of two stages. The simulation results reportedly confirmed that with QC_ALF "on" performance as the reference point, the joint filtering can potentially further improve the coding efficiency by up to 1.5% BD rate reduction in some cases. For Class A tests, the average BD rate reduction were 0.7% and 0.6% in both test cases. This could reportedly imply that performance of QC_ALF can be improved further, particularly for high resolution sequences. The simulation results cover encoder_randomaccess and encoder_lowdelay configuration settings with full length and all classes of the sequences as defined in the Tool Experiment 10 subtest 2 category. The Excel spreadsheet results were provided. This contribution was reportedly not intended to propose an alternative technology to replace QC_ALF in TMuC but rather to be aiming to assist the current general effort on evaluating the existing TMuC video coding tools, such as that in Tool Experiment 12 (TE12) and 10 (TE10). The results reported independently in this document regarding the adaptive loop filtering technology can be treated as additional information to all other related test results in TE12 and TE10.

It was remarked that a gain (0.6-0.7% BR reduction) was achieved only for class A and B partially, and that on average there was practically no gain. Decoder runtime increases by 15-20%.

No action was taken in response.

JCTVC-C211 Inner block oriented ALF processing for memory & bandwidth reduction [Y. Sohn, K.-H. Lee, B.-K. Lee, I.-K. Kim (Samsung)]

This proposal described an adaptive loop filter that was asserted to have low memory, bandwidth and computational complexity requirements. ALF in TMuC requires spatial neighboring reconstructed data for CU based ALF processing. This document proposes applying filtering only to a restricted inner LCU area where there is no need to use data outside of the LCU for filtering. This approach was reported to result in marginal performance degradation with memory, bandwidth and computational complexity reduction. The experimental results reportedly showed 1.0% performance degradation in low delay configuration and 0.8% performance degradation in random access configuration, respectively. In expense of this performance degradation, it was asserted to reduce about 24 Kbytes external DDR memory and 0.77 Kbytes internal SRAM, and to reduce memory bandwidth about 5.98 MHz in the case of 4Kx2K@30p video sequence and 64 bit BUS. It was asserted to also reduce computational complexity for ALF processing itself.

Concern was expressed that for two adjacent filtered blocks, boundary effects may occur.

This would need to be a normative decoder issue (making it switchable would not help to reduce maximum decoder complexity). Most experts believed that not every application case of loop filtering would benefit from that.

A suggestion by the chair was that if limitation of memory accesses is desired, it might be better to limit the number of filtered blocks in a profile.

JCTVC-C219 In-loop reference frame denoising [P. Amon, A. Hutter, E. Wige, A. Kaup (Siemens)]

This proposal presented an algorithm for in-loop denoising of the reference frame. The algorithm modifies the temporal predictor while the decoded picture is unchanged. Knowledge of the noise power within the reference frame is used in order to affect the inter frame prediction. For noise filtering of the reference frame, a denoising algorithm is implemented inside the AVC reference software JM 15.1. It was reported that the bit rate can be decreased for (high resolution) noisy image sequences especially for higher qualities at medium to high data-rates.

The following remarks were recorded:

• 
  The suggestion was to take into account additive noise in the design of the filter coefficients.
  
• 
  The implementation was in the AVC context.
  
• 
  The algorithm currently does not work on intra.
  
• 
  It is not necessary to transmit filter parameters; as these are estimated at the decoder.
  
• 
  The reported improvement was mostly at high bit rates, and may be hardly visible.
  
• 
  The concept may eventually help for applications that require lossless or near-lossless coding (e.g. medical, studio).
  
• 
  The algorithm does not remove quantization noise. It is more intended to separate content that is consistent over time from noise that is uncorrelated between frames and therefore disturbs the prediction. At low rates, it was suggested that this would be removed anyway by the quantization.
  
• 
  Further study was suggested.