JVET-AL0097 AHG 9: Gaussian blur filling method in display rectangles SEI message [C. Kim, H. Tan, J. Nam, J. Lee, J. Lim, S. Kim (LGE)]
It is asserted that the current display rectangles (DR) SEI message has two problems:
- Lacks the signalling of kernel size which is needed for Gaussian blurring operation. The lack of such signalling may lead to inconsistencies in blur effects and computational complexity.
- The blur sigma value is defined as an integer, whereas floating-point representation allows for finer control over blur intensity.
This contribution proposes to introduce kernel size signalling and support floating-point sigma values, improving consistency, flexibility, and accuracy in blur effect.
Item 1 agreed.
It was suggested that an alternative to signalling the numerator and denominator of a floating point number may be using u(v) coding, but that requires normalization in the range of 0 to 1.
A bit shift is also possible, and would avoid the need for a division, such as expressing in units of 1/16ths.
It was noted that the provided semantics are using a / operator for the division, which will clip to the nearest integer, which may not be the intent, which may be to use ÷.
It was noted that sigma equal to 0 doesn’t make sense.
Side activity for the supported range and number of bits for the numerator and denominator and signalling alternatives.
Further discussion on 3 April.
Agreed to keep the blur sigma signalling as u(8) with a minus1 and to modify the semantics to indicate that it is in units of 1/16ths.