Does some noise = better compression

grifter09

Active member
My apologies if this has been discussed already.

Has anybody noticed that the compression seems very blocky at lower iso values, particularly in the blacks.

It appears to me that at iso 640 the blockyness in the blacks gets randomised by the noise. I know a lot of the high end lads here who used the d90 specifically added grain in post before encoding to avoid what im talking about.

If i can show reference to an already posted (and much appreciated, as all tests are) series of tests. look at the manhattan skyline from iso100 up to 640 and watch it improve up to that point. you need a bright screen to see what im talking about.
http://www.dvxuser.com/V6/showthread.php?t=185471&page=4

So, would others agree that one needs noise with this 7d codec we have. ISO640 to be precise?
 
Last edited:
I don't know how that would make any sense, since more noise means more complicated overall image detail to compress. Perhaps what you mean isn't "better compression," but rather having blocking artifacts that are less distracting because they appear more "randomized."
 
well yes, this would be difficult to quantify. but to my eye, i prefer a picture when i dont see a big cluster of the same shade of black, only to be neighboured by a lighter shade bordered by a very mosaic look. And that is what im seeing here with this footage. Until 'some' iso noise is introduced.
 
I haven't tested this with the 7D, but what you are noticing is a well-known phenomenon, dating back to the original Canon XL1.

An 8-bit codec creates banding, because it can only resolve 256 shades or tones. And compression likes to "round things off", if two pixels are very, very close to each other then the compression might decide to make them both be the same shade, thus allowing it to compress them to half their size, etc.

So, the net result is that when you have a gradient, a very very smooth transition from one shade to another, it's likely that you'll see "banding" in that image because the compression will tend to round down the lower shades and round up the upper shades, trying to make them all appear similar so they'll compress easier. Which accentuates the differences and makes two different bands of color or bands of shades.

So, with some noise in the image, the compression can't get away with that trick because there's too much variation. Which means that you preserve the differing shades better. But, you'll pay a price somewhere else in the image because the efficiency the compression gains by making those bands, will be lost in the noisier image. So you might end up with an overall lower-res image, or with mosquito noise or something else.

Way back when in the XL1 days, I think it was Michael Pappas that pointed out the banding issue on the XL1, pointing the camera at an even surface like a wall (that had a light source pointed at it) would show distinct rings of shades. Canon looked into it and came back saying "well, just add 6dB of gain, that'll take care of it". The idea being that the additional noise would mix up the bands and keep the compressor from being able to merge the image all into bands.

The other way to fix it, of course, is to use a 10-bit codec, but that's not an option on these cams.
 
yes thankyou Barry, that is what i am talking about alright. Sometimes i find it hard to articualte or I may not know the exact phraseology. (despite having worked in this business a long time).

So based on this do you think this issue also needs to be considered when shooting or should one simply just accept it & boil it down to 'the way it is at the moment' ?

I know one could probably crush the range of colours this problem appears in when it rears its head in post. But if introducing a small bit of noise eliminates it without ruining the picture, then perhaps this is the way to go.

Im simply in the process of deciding where (in terms of iso value) exactly the tipping point lies in relation this with the 7d footage specifically, and then make a decision as to whether or not the image is acceptable.
 
Last edited:
Visible blocking is caused because of 255 luma values, actually 235-16 = 219 values. This is not enough for image with wide latitude. Gamma correction helps, but even with it and with other tricks you get at most 50:1 contrast ratio WITHOUT BANDING.

Noise works as sort of dithering, it may look visually better, but compression will suffer. It is better to shoot without noise and to add noise later if so wanted.

You can play with gamma and contrast range to get more near-blacks but always at a cost. You will have to give up either midtones or highlights or latitude. This is why AVC-Intra with its standard 10-bit encoding is such a big deal.
 
Back
Top