Test 3: Additive gain of LM and RM
|
|
All Intra HE Main-tier
|
All Intra HE High-tier
|
All Intra HE Super-High-tier
|
|
|
Y
|
U
|
V
|
Y
|
U
|
V
|
Y
|
U
|
V
|
Test 3.1
|
YCbCr 4:4:4
|
−2.4%
|
−9.5%
|
−9.7%
|
−2.7%
|
−6.2%
|
−8.5%
|
−2.7%
|
−4.3%
|
−6.1%
|
|
YCbCr 4:2:2
|
−1.9%
|
−7.0%
|
−6.1%
|
−1.6%
|
−3.6%
|
−3.7%
|
−1.1%
|
−2.1%
|
−2.1%
|
Discussion notes:
-
Was complexity analyzed?
-
Among 1.1 and 1.2 tests (which only apply to intra), the first tested variant performed the best (YUV) domain. The 2nd variant of 1.1 was proposed as a lower complexity scheme than the first variant, but had some boundary sample memory increase and lower performance and had a division problem for the 4:2:2. The 1.2 scheme is also somewhat complex as compared to 1.1. It was remarked that the use of different reference boundary samples in the LM chroma scheme is a complexity issue, and this had not changed. The gain for AI was about 2% for luma and 5% for chroma (about 3/4 of that for 4:2:2, more for 4:4:4).
-
RM scheme does not significantly help beyond LM, so not particularly interesting as a combination.
-
The LM scheme has decoder inference of the scaling parameter estimated at the TU level based on boundary sample values.
-
The CU-adaptive colour transform scheme tested 2.1 has relatively little benefit relative to LM for AI YUV 4:4:4 and does not work properly (as tested) for 4:2:2.
-
It was remarked that the test 2 colour conversion may have had some issue for YCbCr conversion equation calculations.
-
In test 2, per component results varied. It was remarked that the relative quantization fidelity assigned to the three components will affect the results, and the different colour transforms each produce different energy and correlation characteristics. The relative importance of the three components seemed difficult to ascertain. The BD BR curves were suggested to potentially have some overlap-region measurement reliability issues. Ideally it would be desirable to align the R-D curves for two of the three components to examine the effect on the third. Overall, the results seemed very difficult to analyze. Different colour transforms would each have different characteristics (although many would probably be better than RGB). It was remarked that the bit allocation generally assigned about 60–65% of the bits to luma (very similarly) for all of the tested cases, except when coding in RGB – for which each component got roughly the same number of bits.
-
Out of the tested schemes, it seems that LM chroma variant 1 is the most promising. It only applies to intra. However, the luma/chroma cross-pipeline dependencies and reference sample set difference remain troubling.
Plan: Further study in CE for LM chroma, including testing alternative reference sample set and ultra-high bit rates.
Dostları ilə paylaş: |
