Comments on: Adaptive Post/Loop Filters in JM/KTA – Part 2

By: Sandip Ray

Sandip Ray — Mon, 05 Oct 2009 10:02:39 +0000

Adaptive loop filter is an additional block apart from traditional in-loop deblocking filter. Wiener filter is a MSE-optimal stationary linear filter which estimates the original signal from the signal corrupted by additive noise. Here additive noise is the coding noise(mostly due to lossy quantization). Is it fair enough to assume that coding noise of every pixels in a frame have the same statistics ? Or should we group pixels with identical coding noise and apply Wiener filter on pixels within the same group ? I am thinking of some kind of segementation as proposed by Zhang.

By: Wenhao Zhang

Wenhao Zhang — Mon, 28 Sep 2009 07:34:29 +0000

In-loop filter is definitely better than the post filter. Qualcomm has proposed a similar Wiener adaptive filter scheme which used as post filter, but their gain is much smaller compared to BALF.

By: Jie Dong

Jie Dong — Sun, 27 Sep 2009 03:42:44 +0000

Using the filter in loop can improve the capability of predicting other pictures.

By: hurumi

hurumi — Sat, 26 Sep 2009 17:48:53 +0000

Do you think that BALF or QALF can provide much better efficiency compared to post-filter only design? I’m not sure that in-loop type is required for this kind of filtering.

By: Yu Liu

Yu Liu — Wed, 16 Sep 2009 07:42:09 +0000

The proposals of BALF/QALF didn’t describe how to get the array filter_luma_base/filter_chroma_base. I guess that it is obtained by a training set.

By: Yuping Wei

Yuping Wei — Tue, 08 Sep 2009 02:54:20 +0000

Who can tell me how the array filter_luma_base and filter_chroma_base come?

By: Wenhao Zhang

Wenhao Zhang — Mon, 24 Aug 2009 10:15:18 +0000

I think the key point to further improve the performance of loop Wiener filter is video content segmentation. After all, Wiener filter works well only on stationary signal. It is unavoidable that current loop Wiener filter reduces error of many pixels, meanwhile, enlarges error of some other pixels. Some adaptive segmentation scheme is necessary. Brain racking…