Canon 5D III on chip noise reduction?

[uaiomex]

This thread is a bit above me but I wish to contribute with my two cents. So far, all the pictures I've seen from the 5D3 look amazingly clean and sharp. But... when viewed at 100% they look plasticky and smeared, unlike the files of my 5D2.

Perhaps "on chip noise reduction" is Canon'd new little trick that works by canceling the offending pixels by some sort of "cloning" or partial cloning of the noisy-pixel. Some kind of fill-content-aware at the pixel level.

There's got to be an explanation. Not even Canon can screw it so bad. Perhaps Canon criteria is that it doesn't matter if the pics don't look sharp at 100% as long as the pictures look really sharp and amazingly clean on a printed page and at screen size.

Disclaimer: I'm so disappointed for the swivel screen missing and the extra $1K asking price, that perhaps I'm inventing stories to put up with this shit.

Eduardo

[DaveCurtis]

The images over at http://www.imaging-resource.com/ look really impressive at high ISO compared to the MarkII. Very clean and plenty of fine detail.

I think Im in on this baby.

[Ray]

The images over at http://www.imaging-resource.com/ look really impressive at high ISO compared to the MarkII. Very clean and plenty of fine detail.

I think Im in on this baby.

There appears to be something less than a 1 stop DR/Noise improvement to my eyes. Shadows look particularly clean in the 5D3 shots, initially giving the impression that there might be a 2 stop improvement. However, the same 5D3 image with cleaner shadows also shows smeared fine detail elsewhere, compared with the 5D2 image at 1 stop lower ISO.

I get the impression that more aggressive noise reduction has been applied to the 5D3 image, but I admit I haven't downloaded any RAW images yet.

[asf]

You should download the raws then, much more impressive. 

[BJL]

sensel->g1->add noise->g2->ADC

... if g1 has to be limited for low ISO, this means that the analog stages before ADC (g2) allows limited input level before reaching saturation?

That is the way I understand it, with one note. The signal at g1 (on a sense capacitor at the edge of the sensor) is a charge, whereas the input to the ADC is a voltage, so g2 involves charge-to-voltage conversion. This CTV might well be done on the sensor chip itself, since most of the sensor specs I have seen givr output voltages and conversion factors in mV/e- as features of the sensor itself, not its support chips. On the other hand, ISO adjustment seems to typically be done off the sensor chip, which suggests that it is done with a voltage amplifier  off sensor, before ADC. So one likely picture is
Sensel - charge gain g1 - transport along sensor edge - CTV conversion - voltage gain g2 - ADC

Anyway, the possibilities include:
1) The CTV stage having a maximum input charge, less than the charge that would be produced by applying the maximum gain at stage g1 to the charge in a full well.
2) a limit on the current that can be carried from sense capacitor to CTV along the edge of the sensor, which due to the read-out speeds involved sets a maximum on the charge that can be transferred from sense capacitor to the CTV stage.
3) The sense capacitor having a maximum capacity, again less than what would be given by applying maximum g1 to the charge in a full well.

However IIRC, the charge amplification g1 as described by Canon was done by varying the capacitance of the sense capacitor (same charge in sensel delivering same voltage on read circuit which delivers a charge to the sense capacitor proportional to its capacitance), and this seems to rule out (3).

P. S. I overlooked the fact that some ISO speed adjustments are actually done in the digital domain, so the whole picture could be:
Sensel - charge gain g1 - transport along sensor edge - CTV conversion - voltage gain g2 - ADC - digital gain g3.

And in fact the evidence in previous threads is that beyond a quite modest ISO speed (800 or less), there is no benefit to further analog amplification, because the SNR of the signal is already well below what the remaining analog components offer. So it could be that there is no variation in the voltage gain g2, and ISO speed is adjusted only by varying the early charge gain g1 and the digital gain g3 (which is actually a reduction for the special low ISO settings.)

[BJL]

Has anyone compared the dynamic range, via shadow handling at the minimum normal ISO speed of 100?

ISO 50, with special setting "L1", seems to be "pull processing", probably achieved by increasing exposure by one stop but with the same handling of the analog signal and then reducing level by one stop in the digital domain, so sacrificing one stop of highlight headroom above the metered level in exchange for better handling of lower levels. If so, that will not increase DR; it could even reduce it in practice, due to blown highlights, so could just confuse attempts to assess the DR.