Auto ISO and number of RAW bits needed at higher ISO values

These very interesting questions were asked by John in the 50D first impressions thread originally. Since they are rather important to discuss even independently from the 50D, I have moved them here. Cem.

Rachel Foster said:

Sounds good. I considered the 50d very seriously.

I rarely shoot on anything but M these days. Does anyone use the settings like this CA?

Click to expand...

I'd like to shoot in manual Av and Tv, but with ISO floating with the metering, with *every* ISO available, but limitable by the user, and EC setting applied to the ISO (set it to - for more headroom, + for exposing low-contrast and/or high-key scenes, etc). The lack of such a mode in the new Canons pains me. I've waited far too long for this very basic and necessary feature. I don't know why Canon is so adamant about omitting this very valuable feature, which is most closely related to what a photographer really ultimately wants the camera to do; capture a slice of time in a scene with a certain aperture, but with minimum noise, which is what you would get if the camera used the highest ISO that didn't cause clipping.

Ironically, the first adopters of such a mode (called TAv by Pentax) were cameras where the RAW data wasn't significantly cleaner at high ISO than low ISO under-exposed, unlike Canons, which can have as little as 14% as much read noise at ISO 1600 as ISO 100 under-exposed 4 stops, and have been like that for years!

John Sheehy said:

The lack of such a mode in the new Canons pains me. I've waited far too long for this very basic and necessary feature.

Click to expand...

Well, I wrote that as a non-owner of the 50D. Since I wrote that, I have bought a 50D. Lots of nice features over my 30D (sensor cleaning and live view), but now something else pains me. I saw my 8 GB card fill up in a slow day - and I remembered that this new camera is 14-bit. I took a number of shots at ISOs 6400 and 12800. What are these companies thinking? Why do they bloat the RAW data unnecessarily? ISO 12800 needs about 4 or 5 bits, but the camera is writing 14-bit files, with 9 or 10 bits of incompressible noise. ISO 100 probably can't even make good use of the 13th bit. What a waste of space!

I need to teach a RAW math refresher course to the people at Canon. I'd take a "significant bits" file size reduction at full resolution any day over "sRAW".

How about a utility to throw away the unneeded bits?

Asher

John Sheehy said:

...and I remembered that this new camera is 14-bit. I took a number of shots at ISOs 6400 and 12800. What are these companies thinking? Why do they bloat the RAW data unnecessarily? ISO 12800 needs about 4 or 5 bits, but the camera is writing 14-bit files, with 9 or 10 bits of incompressible noise. ISO 100 probably can't even make good use of the 13th bit. What a waste of space!

I need to teach a RAW math refresher course to the people at Canon. I'd take a "significant bits" file size reduction at full resolution any day over "sRAW".

Click to expand...

Asher Kelman said:

How about a utility to throw away the unneeded bits?
Asher

Click to expand...

Hi Asher,

There is one, save the pictures in GIF format instead of RAW, LOL. Now you must sure know how your photos look like if you accidentally save them in GIF using only 4 bits, don't you?

Joking aside, I'd like to think this through. IMO, there is no possible way for the camera to know that maybe 9-10 bits are "incompressible noise" so it should be thrown away. If it could, then it'd use that knowledge to eliminate the noise to start with. So let's suppose that we can write intelligent algorithms to deal with this. My guess would then be that in the current state of the released technology this cannot be executed practically. Unless one would be willing to live with frame rates of 0.1 fps or some other downsides which I cannot foresee right now. I am sure that we'd have to wait for a DIGIC XII or so which will have enough processing power.

Besides, given the prices of memory, I'd prefer to pay $60 for an extra 8GB CF card than I'd pay $600 extra for a camera which is capable of bit reduction so that the images are smaller. Economically, it doesn't make sense to me.

Wrt to the 13th bit in ISO 100, I am empirically convinced that it helps. I have seen time after time that that I can open up the shadows more during RAW conversion (thanks to the extra bits) for my 40D images compared to the images coming from my 5D.

Just my Euro 0.02

John Sheehy said:

Why do they bloat the RAW data unnecessarily? ISO 12800 needs about 4 or 5 bits, but the camera is writing 14-bit files, with 9 or 10 bits of incompressible noise. ISO 100 probably can't even make good use of the 13th bit. What a waste of space!

Click to expand...

This is why they invented a special file format called a Jpeg!

This is also why I take 60GB of CF cards to a shoot!

Will Thompson said:

This is why they invented a special file format called a Jpeg!

Click to expand...

JPEG and RAW are two completely different things. A JPEG is dressed up to look presentable in final display; it is not optimized to be a container of the initial RAW capture.

My assessment of the needed bits in a RAW file is not some whim; I have done quite a bit of testing, and there really is no benefit in having more levels in a RAW file than is necessary to have the standard deviation of read noise and the dark current noise combined use more than about 1.3 to 1.4 levels. The minimum noise occurring in the cleanest 14-bit DSLRs at ISO 100 is about 4.9 levels, and the cleanest at ISO 1600 about 12.8 (from the Canon 1D3). ISO 12800 on all cameras has about 8x the read noise of ISO 1600, or about 100 levels, or 70x as much as necessary, in the best case.

Of course, for anyone doing stacking of RAW images such as for astrophotography, the extra digitization is more useful, because the small amount of extra signal keeps adding, while the noise decimates itself with stacking.

Cem_Usakligil said:

Hi Asher,

There is one, save the pictures in GIF format instead of RAW, LOL.

Click to expand...

History shows that those who get laughed at are often in good company.

Now you must sure know how your photos look like if you accidentally save them in GIF using only 4 bits, don't you?

Click to expand...

We all have a picture about what happens with a nice, smoothly graded image. The days of saving as a 16 or 256 color GIF, though, were the days of images generated by computers, or sampled with better lighting than DSLRs at ISOs in the thousands. Then you also have the issue of the fact that the demosaicing of RAW data creates extra levels.

Joking aside, I'd like to think this through. IMO, there is no possible way for the camera to know that maybe 9-10 bits are "incompressible noise" so it should be thrown away. If it could, then it'd use that knowledge to eliminate the noise to start with.

Click to expand...

No; you do not reduce noise by quantizing. You always increase noise by quantizing. The question is at what level of quantization is it, for all practical purposes, not a loss. My assessment is that 1.4 levels to match the standard deviation of the read noise is ample.

So let's suppose that we can write intelligent algorithms to deal with this. My guess would then be that in the current state of the released technology this cannot be executed practically. Unless one would be willing to live with frame rates of 0.1 fps or some other downsides which I cannot foresee right now. I am sure that we'd have to wait for a DIGIC XII or so which will have enough processing power.

Click to expand...

All that has to be done is to bit-shift the digitized values while writing to the card. No computer operation could be much simpler or faster.

The number of useful bits is fixed. Cameras of the same model do not vary significantly in read noise. Any extra noise near black for long exposures would only reduce the needed precision, so it can be ignored for this purpose.

Besides, given the prices of memory, I'd prefer to pay $60 for an extra 8GB CF card than I'd pay $600 extra for a camera which is capable of bit reduction so that the images are smaller. Economically, it doesn't make sense to me.

Click to expand...

It would cost literally nothing. You are imagining a complexity which does not exist. Simply dropping unnecessary bits is all that is required. There is no need to dither, as the read noise does all the dithering you need, and that is the very thing that makes it useless to have the extra bits.

Wrt to the 13th bit in ISO 100, I am empirically convinced that it helps. I have seen time after time that that I can open up the shadows more during RAW conversion (thanks to the extra bits) for my 40D images compared to the images coming from my 5D.

Click to expand...

The 40D has about 1/2 stop less read noise than the 5D at the pixel level, at ISO 100. It still has 1/2 stop if you quantize the 40D RAW data to 12 bits.

Besides, an improvement could exist just because the 14 bits increases the *mandatory* precision of a RAW converter. A converter could use that extra level of precision anyway, or, the RAW file format could just decompress to an RAW with padded zeros to force the precision in the converter, but not in the storage of the RAW data.

Asher Kelman said:

How about a utility to throw away the unneeded bits?

Asher

Click to expand...

That would be easy for DNG, which is well documented. For proprietary RAWs, it would have to decompress and recompress the data. In either case, you would have have to pad with zeros instead of bit-shifting the data, and whether or not that automatically makes the compressed files smaller, without a new codec, I don't know off the top of my head for DNG or any other format.

John Sheehy said:

JPEG and RAW are two completely different things.

Click to expand...

I never said to use a Jpeg as a replacement for a RAW. A Jpeg is available for those who complain that cards fill up to fast.

Will Thompson said:

I never said to use a Jpeg as a replacement for a RAW. A Jpeg is available for those who complain that cards fill up to fast.

Click to expand...

But then JPEG would replace RAW. You can't replace it and not replace it.

NO! It is a choice not a replacement.

John Sheehy said:

But then JPEG would replace RAW. You can't replace it and not replace it.

Click to expand...

Will Thompson said:

NO! It is a choice not a replacement.

Click to expand...

Guys, guys, you are both right, kind of....

It is a choice indeed to either use RAW or JPG due to the file size economy.
But when one chooses to have JPG instead of RAW, you don't have RAW anymore.

PS: JPG instead of RAW is in my opinion a choice only from the lower file size point of view. JPG can never replace RAW from a functionality point of view.

John Sheehy said:

...No; you do not reduce noise by quantizing. You always increase noise by quantizing. The question is at what level of quantization is it, for all practical purposes, not a loss. My assessment is that 1.4 levels to match the standard deviation of the read noise is ample.
...
All that has to be done is to bit-shift the digitized values while writing to the card. No computer operation could be much simpler or faster.
...
The number of useful bits is fixed. Cameras of the same model do not vary significantly in read noise. Any extra noise near black for long exposures would only reduce the needed precision, so it can be ignored for this purpose.
...
It would cost literally nothing. You are imagining a complexity which does not exist. Simply dropping unnecessary bits is all that is required. There is no need to dither, as the read noise does all the dithering you need, and that is the very thing that makes it useless to have the extra bits.


....

Click to expand...

Hi John,

I am in a rush now so I'll try to come back to this later today. But let me understand this properly. Are you talking about getting rid of the "least significant bit(s)" in the RAW file since they are below the noise treshold and hence are insignificant to be recorded? We would then just leave the least significant bits zero and either compress them or not record them at all (but set a flag somewhere in the RAW indicating that this is so). Is this what you are referring to? Or am I totally off the mark?

Cheers,

Cem_Usakligil said:

Hi John,

I am in a rush now so I'll try to come back to this later today. But let me understand this properly. Are you talking about getting rid of the "least significant bit(s)" in the RAW file since they are below the noise treshold and hence are insignificant to be recorded? We would then just leave the least significant bits zero and either compress them or not record them at all (but set a flag somewhere in the RAW indicating that this is so). Is this what you are referring to? Or am I totally off the mark?

Click to expand...

That's pretty much what I'm saying, and also that the conversion process should always work with greater internal precision than the significant bits.

John Sheehy said:

Well, I wrote that as a non-owner of the 50D. Since I wrote that, I have bought a 50D. Lots of nice features over my 30D (sensor cleaning and live view), but now something else pains me.

Click to expand...

Hi John,

Congratulations on your new 50D. I'm looking forward to your analysis of it's Raw quality performance. One needs to spend some time with a camera to do the in-depth data gathering, I know from personal experience.

I saw my 8 GB card fill up in a slow day - and I remembered that this new camera is 14-bit. I took a number of shots at ISOs 6400 and 12800. What are these companies thinking? Why do they bloat the RAW data unnecessarily? ISO 12800 needs about 4 or 5 bits, but the camera is writing 14-bit files, with 9 or 10 bits of incompressible noise. ISO 100 probably can't even make good use of the 13th bit. What a waste of space!

Click to expand...

While I understand your observations, I've been shooting my 21Mp 14-bit 1Ds3 for a while now, and a decent mobile backup harddrive (Hyperdrive Colorspace) gets more use than ever before. On the other hand, you know better than most that in camera High ISOs aren't necessarily better than lower ISOs, once one exceeds a certain level. And to avoid unfortunate upstream decisions, I'd rather have access to the "Raw deal", and decide myself which bits to drop and how. Besides, averaging noisy images, like in astro-photography, will recover amazing amounts of detail from seemingly random noise.

I need to teach a RAW math refresher course to the people at Canon. I'd take a "significant bits" file size reduction at full resolution any day over "sRAW".

Click to expand...

I wouldn't be surprised if they already know, but still make the decisions based on what will market better, rather than what IS better. The Wedding photography market is presumably a very big user of sRAW files.

Bart