Let’s talk about those Canon EOS RS/R5 specifications

Michael Clark

Now we see through a glass, darkly...
Apr 5, 2016
1,781
985
CR guy, I think what paints everyone's expectations is that any new FF camera -- if real -- is going to fall into a 1-series / 5-series / 6-series sort of slot in the lineup, with its relative prestige and asking price, etc. And the spec list we've seen (regardless of who it is aimed at in the market) is comically above/beyond what Canon's brand-spanking-new 1DX3* can do.

That -- above all else -- doesn't compute for many of us, myself included.
  • Tell us it's 45 x 10 and it's the next 5D5, it would be plausible.
  • Tell us it's a gripped 1-series mirrorless with IBIS and 24 x 20, it would entirely be plausible.
  • Tell us it's as-the-rumor-specs-it and it will be a new series even more pricey/prestigious than the 1-series (a 0-series?!) and it becomes I guess vaguely more plausible. (Seems like a $10k body to me, though.)
But as spec'd. No. Still not buying this rumor.

- A

*I recognize a 1-series is much more than pure horsepower specs, and we've seen 5-series outdo 1-series on the odd spec or new feature in the past ...but this rumored list is simply ridiculous.

As far as stills via the OVF goes, thinks about this. The new OVF based AF sensor for the 1D X Mark III is essentially another CMOS sensor, rather than a PDAF line sensor as every DSLR in existence from any company thus far has used. That has to have a higher processing overhead than the older line type sensors. The 1D X Mark III is processing the information from the (dual pixel monochrome?) CMOS AF sensor, the CMOS RGB+IR metering sensor, and then integrating all of that data together, and comparing the results with what is likely the most extensive library of different scenarios ever included in a camera's internal database - in addition to processing the images coming off the 20MP imaging sensor.
 

hne

Gear limits your creativity
Jan 8, 2016
307
25
As far as stills via the OVF goes, thinks about this. The new OVF based AF sensor for the 1D X Mark III is essentially another CMOS sensor, rather than a PDAF line sensor as every DSLR in existence from any company thus far has used. That has to have a higher processing overhead than the older line type sensors. The 1D X Mark III is processing the information from the (dual pixel monochrome?) CMOS AF sensor, the CMOS RGB+IR metering sensor, and then integrating all of that data together, and comparing the results with what is likely the most extensive library of different scenarios ever included in a camera's internal database - in addition to processing the images coming off the 20MP imaging sensor.
The AE sensor data usable for subject recognition is only about 800x500 pixels, or according to Canon "approximately 400'000 pixels". The new AF sensor has 100x more pixels than the 61-point sensors previously used (which have ~70 line sensors). So... another 1Mpx. In total, the amount of data processed is roughly the same as a 1080p video stream. Smartphones have been able to run that kind of data through deep learning neural networks in realtime on general-purpose CPUs for years. But in the last few years we've also seen the appearance of TPUs. I believe there is a TPU core stuck in the DIGIC X chip, because of some mentions of handover to that chip for "AF Priority (people)" in cases where face detection fails.

Deep learning is computationally expensive only in the learning phase. The use of the trained network for classification is definitely something that can be done in embedded systems like a DSLR. There is no database needed in the camera. It has already been reduced to weights:
 
  • Like
Reactions: SecureGSM