Organisation internationale de normalisation

17.3Source Video Test Material

17.3.1.1.1.1.1.1.1JCTVC-D034 Specification of cropped Super Hi-Vision test sequences [Atsuro Ichigaya, Kazuhisa Iguchi, Shinichi Sakaida, Yoshiaki Shishikui]

The test material ad hoc group has discussed new test materials for JCT-VC activity. It has been suggested that the variety of current test materials is not sufficient. This document contributes some cropped Super Hi-Vision (SHV) sequences as new test materials for beyond HDTV sequences. NHK has previously contributed SHV format test sequences, which are in 7680 x 4680/60 fps/10 bpp format, to JCT-VC (JCTVC-A023). Moreover, JCTVC-C055 reported that these cropped sequences, which were in the 2560 x 1600/60 fps/8 bpp format, were available for evaluating new video coding methods and their performance by TMuC 0.7 (JCTVC-C055). In the Guangzhou meeting, a viewing session was held at which the sequences were viewed. This document described the cropped SHV sequences. The contributor proposed these cropped sequences to become a part of Class A sequences for group testing purposes.

The two sequences are called "Nebuta" and "Steam Locomotive Train" (SL), of length 300 frames (5 sec) each, plus one additional frame appended at the end containing a copyright notice (and the coding process for experiments should not include that frame).

Coding PSNR results for AI HE, RA HE, AI LC, and RA LC were provided.

It was remarked that these sequences were produced using upsampling from a Bayer-like RGBG capture that was not full resolution for all color components. It was suggested that, in the longer term, true 8K sequences (not upsampled) would be desirable to obtain to avoid potential effects from the upsampling filter.

The original sequences were 10 bits per sample. After the upsampling, fixed rounding was applied to convert to 8 bits per sample.

We currently have two Class A sequences in our common test conditions. Adding these would bring that to a total of four.

It was noted that the lower number of sequences in Class A relative to other classes has, up to now, resulted in Class A being under-represented in overall averages.

It was remarked that the Nebuta sequence contains a significant amount of noise.

It was remarked that the HE configuration actually processes video using 10 bits per sample. However, it does not operate in a 4:4:4 mode.

The contributor can provide the 10 bit per sample sequences.

Planar formatting for the input and output files (probably 16 b per sample, aligned to MSB or LSB? Which endian? – check the JM) was expected to be used to accommodate the 4:2:0 nature.

The usual Hannover ftp site should work for distribution.

Decision: These sequences are adopted into our Class A common test conditions (as additional sequences not replacing either of the two prior ones), using the 10 b version for the HE case. These sequences were uploaded to the BBC server during the meeting, and 8b sequences had previously been uploaded to the Hannover server. The BBC server has also been used for reference encoding bitstreams.

Participants may contact the JCT-VC or test material AHG chairs to obtain access credentials for the sequences.

17.3.1.1.1.1.1.1.2JCTVC-D092 Study on coding performance of HM1 with new test sequences [Teruhiko Suzuki, Ali Tabatabai]

As JCTVC-C168 advocated, it was asserted to be important to test a greater variety of test materials during the development of HEVC. The current test sequences were reported to be similar to each other in some characteristics. For example, a lot of panning sequences are included. It was suggested that new test materials to cover a larger variety of scenes should be collected. This contribution reported the coding performance of existing test materials of VCEG/MPEG. The coding performance was compared with AVC High profile by JM 17.2.

ParkJoy, Crowd Run, and (to a somewhat lesser extent) Night were noted to be especially difficult. It was noted that the graphs in the contribution (v2) had some labeling problems. It was noted that the chroma performance is sometimes not good relative to the JM, and perhaps we should study the default balance between luma and chroma fidelity and perhaps align that more closely to the AVC standard.

It was proposed to include Crowd Run in class B and to include Night and Harbor in Class E (or create a new Class F).

However, the group thought that it may be best not to take that action (at least at this time). Some specific aspects of the discussion are noted below.

It was remarked that Crowd Run (and other SVT sequences such as Park Joy) have very high noise due to being captured with an early capture process (involving conversion from 50 Hz film capture).

It was remarked that Night and Harbor were compressed in the original (using D5) and have some (visually minor but possibly relevant) artifacts as a result.

It was remarked that we should not select sequences just because they provide worst-case results relative to AVC. In fact, it would be undesirable to do so, as this would produce a test set that under-represented the coding capability of HEVC for typical use.

"People on Street" in Class A (and the new Nebuta) was noted to also be noisy.

It was remarked that VQEG has a large number of new HD sequences (and are producing additional ones), and investigation of the use of these may be beneficial.

It was suggested that the Test Material AHG should be continued.

17.3.1.1.1.1.1.1.3JCTVC-D252 Test Material for Screen Content coding [W. Ding, Y. Shi, B. Yin (BJUT)]

This contribution described and introduced a set of test sequences for screen content which were offered to be made available to JCT-VC and parent bodies by Beijing University of Technology. BJUT proposed to adopt these sequences for developing and testing JCT-VC standards. These were provided in YUV 4:2:0 format. They are further described in the tables below: