Organisation internationale de normalisation

11.2.3Subset 3: Image clipping and offset

11.2.3.1.1.1.1.1.1JCTVC-D122 CE8 Subtest3: Picture Quadtree Adaptive Offset [Chih-Ming Fu, Ching-Yeh Chen, Yu-Wen Huang, Shawmin Lei (MediaTek)]

This contribution described a proposal using an adaptive offset (AO). The proposed picture quadtree adaptive offset (PQAO) uses local adaptation to increase coding efficiency. Each picture is proposed to be divided into multi-level quadtree partitions, and the samples in each leaf partition are affected by a band offset (BO) or edge offset (EO).

Both BO and EO classify samples of a partition into groups, and one offset is derived for each group.

The proposal adds another processing stage between the deblocking/debanding filter and the ALF (assuming both of these are active). This additional stage uses its own additional quadtree segmentation map.

For the "band offset", the "band" is a range of sample values, and the five MSBs are an index to determine the value of the offset, so that the offset may be different for each of 32 segments of the range from 0 to 255.

For the "edge offset", each sample is compared to the values of its neighbors to categorize each sample into one of 5 or 7 categories, and each category has an offset value.

When compared with the JCTVC-C500 anchor, PQAO reportedly achieves 1.5% and 2.2% bit rate reductions for HE-RA and HE-LD, respectively. For HE, the encoding time is increased by about 1%, and the decoding time is increased by about 2%.

For LC comparisons, the runtime increase for the decoder was more substantial.

Subjective testing was done in Guangzhou, with no significant difference observed.

The processing is parallelizable – output samples do not become input to the processing of other samples.

The processing crosses CU boundaries, so if processing CU-by-CU, it is necessary to delay the processing until the CU to the right and the CU below the current CU is available.

It was noted that ALF also has some offset processing in it.

It was remarked by several participants that ALF has some similar types of processing in it that have not been fully analyzed, and it would be beneficial to try to find a way to find gain by harmonizing the gains found in this with (or into) ALF or otherwise improving ALF rather than adding another processing stage.

It was remarked that JCTVC-D119 is an effort in that direction.

11.2.3.1.1.1.1.1.2JCTVC-D183 CE8 Subset3: Cross verification on Picture Quadtree Adaptive Offset (JCTVC-D122) [I.-K. Kim, T. Lee (Samsung)]

Cross-check of JCTVC-D122. The source code was checked and compiled, and found matching results.

11.2.3.1.1.1.1.1.3JCTVC-D123 CE8 Subtest3: Controlled Clipping [Yu-Lin Chang, Chih-Ming Fu, Yu-Wen Huang, Shawmin Lei (MediaTek)]

This contribution described a proposal on controlled clipping (CC). The main concept of CC is to signal the minimum and maximum of original pixel values for predicted or reconstructed pixels to be clipped within the range between the minimum and the maximum. In this way, the distortion between the original video and the decoded video can reportedly be reduced. The proponent applied CC at four different positions of the coding path, i.e. post-prediction, post-reconstruction, post-deblocking, and post-adaptive loop filtering (ALF). Moreover, CC also can be applied as a post processing outside of the entire decoding process and prediction loop. The in-loop CC can be applied at each of the four positions, while the post-loop CC is applied at the output of the decoder. The output pixels of the post-loop CC are not stored in the reference picture buffer but only used for display. Different in-loop/post-loop CC options are tested, and it was reported that the post-loop CC (but with CC syntax) can achieve the best coding efficiency. Comparing with the anchor in JCTVC-C500, the post-loop CC could reportedly achieve 0.6% and 0.4% bit rate reductions for high efficiency random access (HE-RA) configuration and high efficiency low delay (HE-LD) configuration, using CC information sent in the slice header. When the CC information is sent in another NAL unit rather than, the bit rate reductions dropped to 0.4% and 0.2% for HE-RA and HE-LD, respectively, due to the increased overhead. There was reportedly almost no increase in encoding or decoding time.

It was asked how the encoder decides what clipping range to indicate. In the tested method, the encoder measured the minimum and maximum sample values of the source video.

The proposal suggested having a clipping range per component, with syntax supporting either sending three ranges (Y, Cb, Cr) or two ranges (Y and chroma) or just a "broadcast legal" flag, or predictive coding from picture to picture.

It was asked whether ranges in the RGB domain might be better to signal than in the YUV domain, as conversion to RGB is common for display.

It was remarked that, at least when the clipping range is just the "broadcast legal" range, it is likely to already be understood that values outside the range are interpreted as visually clipped to black and white, and the PSNR measure is not really producing any benefit to the interpretation of the values.

It was remarked that if this is operating out of the loop, if it is desirable, it could be handled as something the decoder is not required to do (e.g., using an SEI message).

It was remarked that cascaded encoding followed by decoding followed by re-encoding with clipping between the two encoding stages might actually degrade the signal, as some "headroom" and "footroom" may actually be desired in the signal flow chain. Moreover, the clipping could result in failing to detect, at a destination, that the signal characteristics are mal-adjusted at the source.

Some participants characterized this technique as perhaps an artificial PSNR boosting measure in a manner that is not likely to actually provide a visual picture quality benefit.

11.2.3.1.1.1.1.1.4JCTVC-D077 CE8 Subset3: Verification results of MediaTek’s Proposal (JCTVC-D123) [I. S. Chong, M. Karczewicz (Qualcomm)]

Cross-check of JCTVC-D123. The results reportedly matched those reported by MediaTek.