I'm a programmer in a very different field, and I'm just starting to try to get up to speed with understanding the ML / Digic context, so please excuse any naivety in this question:
One way that ArriRaw, amongst others, reduces its data requirements is to write the 16bit linear data from the sensor into an uncompressed raw 12bit log data stream, which - given the familiar inefficiency of linear encoding of light values - drastically reduces frame sizes.
The current 14-to- 12- and 10-bit strategy is an amazingly impressive piece of work, and I can see took years to achieve, and a lot of specialist knowledge. If I understand correctly, it's a bit-chop lossy reduction in representation precision. Obviously the raw-log-encoded method would be less of a quality drop from 14- to 10-bit, if it existed.
So before I 'go off on one' (as my old AI programming tutor used to put it when he saw me disappearing down a research rabbit hole / dead end again):
Is e.g. a 14-bit to 10-bit log-encoding of the raw uncompressed (linear) sensor data - rather than a lossy bit-chop - computationally plausible in this context? Or is it waaay too ambitious for the resources available on-camera during recording?
It's caught my imagination, but I realise it might be a silly idea if you know enough - so if that's the case, I'll be sensible and take on a far more manageable task from the 'to-do' list as a 'teach myself' project instead...
Thank you!
One way that ArriRaw, amongst others, reduces its data requirements is to write the 16bit linear data from the sensor into an uncompressed raw 12bit log data stream, which - given the familiar inefficiency of linear encoding of light values - drastically reduces frame sizes.
The current 14-to- 12- and 10-bit strategy is an amazingly impressive piece of work, and I can see took years to achieve, and a lot of specialist knowledge. If I understand correctly, it's a bit-chop lossy reduction in representation precision. Obviously the raw-log-encoded method would be less of a quality drop from 14- to 10-bit, if it existed.
So before I 'go off on one' (as my old AI programming tutor used to put it when he saw me disappearing down a research rabbit hole / dead end again):
Is e.g. a 14-bit to 10-bit log-encoding of the raw uncompressed (linear) sensor data - rather than a lossy bit-chop - computationally plausible in this context? Or is it waaay too ambitious for the resources available on-camera during recording?
It's caught my imagination, but I realise it might be a silly idea if you know enough - so if that's the case, I'll be sensible and take on a far more manageable task from the 'to-do' list as a 'teach myself' project instead...
Thank you!