All proposals increase complexity to some extent
-
Adding additional transform bases (DST and/or non-TM DCT)
-
Matrix multiplication (no fast algorithms tested in most cases)
-
Additional quantization matrices
-
Additional scan orders
Rather than investigating runtimes, for cases of transforms, paper analysis of complexity would be desirable (or separate recording of runtime for the transform/quantization/scanning parts).
From the results shown, it is not clear yet whether all the complexity-increasing building blocks are necessary, or how much benefit would remain if some of these were removed.
It would be desirable to not use multiple transforms, or to restrict the number of block sizes for additional transforms to as few as possible.
It will be necessary to look into combination of improved intra prediction vs. simple transforms
Further study in CE – one major goal of CE’s should be to investigate the additional complexity (not only by runtime measurements, which can be unreliable and can swamp out the effect of a tested technical design element with side-effects and the processing for other components) vs. compression benefit.
Dostları ilə paylaş: | 
