International organisation for standardisation organisation internationale de normalisation

5.20.3Lossless compression

(The relevant AHG report was presented in this context. It recommended to review all contributions.)

5.20.3.1.1.1.1.1.1JCTVC-H0083 AHG19: Method of frame-based lossless coding mode for HEVC [M. Zhou (TI)]

Efficient lossless coding was asserted to be required for real-world applications such as automotive vision, video conferencing and long-distance education. This contribution proposed to install a high-level flag "bypass_coding_mode_enabled_flag" in the PPS to enable a lossless coding mode by bypassing inverse quantization, inverse transform, deblocking filter, SAO and ALF, and to use sample-based angular intra prediction (SAP) in the lossless coding mode for better coding efficiency. The proposed sample-based prediction is the same as the HM5.0 block-based angular prediction in terms of prediction angles and sample interpolation, requiring no syntax or semantics changes, but differing in the decoding process in terms of reference sample selection. In the proposed method, a sample to be predicted uses its direct neighbouring samples for better intra prediction accuracy. Compared to the HM5.0 anchor lossless method which bypasses transform, quantization, de-blocking filter, SAO and ALF, the proposed method reportedly provides an average gain of 8.30% in AI-HE, 8.53% in AI-LC, 2.85% in RA-HE, 2.93% in RA-LC, 2.0% in LB-HE and 2.06% in LB-LC. For class F sequences only, the average gain is reportedly 12.23% in AI-HE, 12.55% in AI-LC, 7.11% in RA-HE, 7.28% in RA-LC, 5.72% in LB-HE and 5.84% in LB-LC. The SAP is reportedly fully parallelized on the encoder side, and can be executed at a speed of one row or one column per cycle on the decoder side.

The contribution proposed sample based prediction (SAP) using same prediction directions as currently in the HM but with sample-by-sample prediction (as it is lossless, it was asserted that this does not matter, and everything can still be LCU-based).

Compared to an HM anchor with transform and quantizer bypass, the gain is 8.3% average for AI (including class F, which shows higher gain than other classes), 2.9% for RA, 2.0% for LD

It was remarked that the HM anchor (without SAP) is the true "bypass mode", which would come practically for free, whereas SAP would require implementation of another building block which might not be desirable in a main profile where many applications might not use it.

Compression ratios were reported as follows: