JVET-AL0123 AHG9: On the encoder optimization information (EOI) SEI message [J. Xu, Y.-K. Wang (Bytedance)]
Chaired by S. Deshpande on 27 March 2025 at 13:25
This contribution proposes the following changes for the encoder optimization information (EOI) SEI message:
- Specify a valid range for eoi_type.
See the notes under item 1 of JVET-AL0056.
- Signal original source picture width minus 1 and original source picture height minus 1.
- Specify a constraint on the original source picture width and height to avoid conflicts in semantics.
It was commented that this is to avoid a case where encoder does something redundant and it does not seem harmful. The exact wording regarding meaning of variable EoiSpatialResamplingFlag was discussed and was asserted to indicate that resampling had taken place. It was commented that may be this can be described in a note with a slightly softer language, also the use of “cropped decoded picture width/ height” seems to refer to codec characteristics. There was some confusion if the proposed language covers both cases. It was asked if we should have interface variables defined to make the constraint very specific. Editors to handle the exact language.
- Clarify the semantics of eoi_object_based_idc.
It was commented by two participants that any optimization done by EOI can be prior to encoding or during encoding. Also encoder implementation may not be clear about what is before and during encoding. It was thus agreed to remove the words “prior to encoding” from bitmask 0x01 semantics.
- Specify a constraint on eoi_object_based_idc to disallow some conflicts.
It was commented that adding it for ( eoi_object_based_idc & 0x01) combination is not necessary. Constraint for the values ( eoi_object_based_idc & 0x04 ) and ( eoi_object_based_idc & 0x08 ) makes sense but it was asked if this should be a note.
Decision (ed.): Delegated to the editors to add a note as describe above.
- Clarify the semantics of eoi_privacy_protection_method_idc.
It was commented that “privacy information” is not be better than “personal information”. The change about “prior to encoding” should not be done as discussed related to item 4 above.
No action on this aspect except change the word “undiscernible” to “unidentifiable”.
- Change the values of different privacy protection methods.
There were use cases described to allow combinations, so no action was taken.
- Fix the semantics of eoi_src_pic_flag.
See the notes under item 3 of JVET-AL0056.
- Enable signalling of eoi_src_pic_flag when eoi_num_int_pic is equal to 0.
It was commented that for the value 0 for eoi_num_int_pics the eoi_src_pic_flag may not be meaningful. It was asked if encoder always knows which is the original source picture if temporal interpolation was done before encoder. It was answered that in that case eoi_num_int_pics can be set to 0 and current if condition does not need to signal eoi_src_pic_flag. A participant thought that the encoder should still know in this case which is the original source picture. It was commented that in general the encoder might or might not know instead of does or does not know.
It was delegated to the editors to clarify the meaning of the semantics of eoi_num_int_pics equal to 0 to clarify that encoder might not know which pictures are the source pictures.
- Signal two flags to indicate horizontal and vertical spatial resampling type.
JVET-AL0310 is related and provides another option and was discussed together with this.
In JVET-AL0310 instead of using 2 flags, 2 idcs are used providing more information.
First question is if we want to allow indicating different resampling factors in horizontal and vertical dimensions and what are the use cases for that. It was questioned if there are practical use cases that picture is stretched in one spatial dimension and contracted or kept the same in the other dimension.
It was commented that this is a bit of a feature change and it was agreed by a proponent that this is not a bug-fix. It was asked if this can be done after v4 and it was commented that the syntax for that would be a bit different. It was commented by proponents that this can be considered a necessary change.
Decision: agreed to implement as per JVET-AL0310.
- Extend eoi_privacy_info_type by 1 bit.
It was commented that the “object” as proposed here is very general. It was also asked if we should only define specific types of objects (e.g. computer screen). It was commented if this would make people not specify type of privacy values defined by just setting this bit to 1 instead.
It was asked if background blurring would set this bit to 1. It was commented that the current bit 0x04 could be used for that.
No action taken at this meeting.
- Specify range constraint for eoi_quant_threshold_delta.
See the notes under item 2 of JVET-AL0056.
Decision: Agreed on items 2, 3 and other items as noted above.
JVET-AL0123 AHG9: On the encoder optimization information (EOI) SEI message