JVET-AD0078 AHG9: On comments for the draft of VSEI [T. Chujoh, Y. Yasugi, T. Ikai (Sharp)]

JVET-AD0078 items 1 to 3 belong to this category.

JVET-AD0078 AHG9: On comments for the draft of VSEI [T. Chujoh, Y. Yasugi, T. Ikai (Sharp)]

JVET-AD0078 items 4 and 5 belong to this category.

This contribution reports two editorial problems and three technical problems in JVET-AC2032, the latest draft of the versatile supplemental enhancement information messages for coded video bitstreams (VSEI) standard (Rec. ITU-T H.274 | ISO/IEC 23002-7) and proposes solutions. One technical suggestion regarding neural-network post-filter characteristics SEI messaged is that when only chroma samples are processed, the integer variable, strengthControlScaledVal, should be derived from the chroma variable, inpTensorBitDepthC. The others technical suggestions are about the post-filter hint SEI message. The equation to derive the syntax elements filter_hint_value should be improved because the normalization value may be too large, and pixel outside of the picture may be indicated. The latter problems concern not only VSEI but also HEVC.

Item 4: There is a technical problem in the equation (95) of subsection 8.30.4. The value of normalization, (2^8+bitDepth-1)² in the following equation (95) may not be correct.

Item 5: There is a technical problem in the equation (95) of subsection 8.30.4. The indexes for the decoded signal s′ may indicate samples outside of the picture.

Item 4: Comparing to the original proposal (which was intended for 8 bit), the normalization value should normalize relative to a value of 1 (scaled by 2¹⁶). The suggested correction formula was “mentally crosschecked” for the cases of 10 bit and 12 bit:

• In the 10-bit case, two 10-bit numbers are multiplied, which gives a 20-bit number, to be normalized by 16 to provide a 16-bit number filter_hint_value
• In the 10-bit case, two 12-bit numbers are multiplied, which gives a 24-bit number, to be normalized by 256 to provide a 16-bit number filter_hint_value
• etc.

The same problem is existing in AVC and HEVC.

Item 5: Due to the fact that offset_x/offset_y are only half of the maximum of cx/cy, respectively, samples outside of the reconstructed picture would be addressed in the formula. Even though the formula comes in a note and is not normative, this is confusing. The proposed solution of clipping to the picture width/height resolves the problem.

Decision: It is agreed that items 4 and 5 shall be resolved by the solutions proposed in JVET-AD0078. This fix should be included in VSEI v3, and in a new version of JVET-AD1004. Furthermore, it should be implemented in the HEVC FDIS 5^th edition. As the FDIS had not yet been submitted for ballot, it was agreed later to leave it to the discretion of editors to make this change.