@bainzy - well, one point in Frank’s approach is to utilize the maximum dynamic range of each color channel during capture. Given the dynamic range a typical color-reversal film is able to display, this is probably a sensible approach.
For cinema type of movies, you probably can set the illumination levels for a choosen reference frame and digitize the whole film with these settings. This is different with the material I am mostly concerned with, Super-8 home movies. Here, there will be over- and, more often, under-exposed scenes interwoven with normally exposed ones. In this case, you will need to readjust illumination intensities on the fly, preferably by some feedback system analyzing the capture in real-time.
Now, in any case, each raw scan will initially not be properly white-balanced; you will need to correct that during post-production. If you are planning to color-grade the final movie anyway, that is probably not a big deal.
In contrast, I wanted a mostly automated system which can tolerate some exposure variation and which would avoid the processing of 14-bit or more raw camera data (with my cameras, I can get anyway only 12-bit raw data). After a lot of experiments, I finally settled for employing exposure fusion, based on 5 images taken ranging over 5 EVs.
What is the difference compared to capture a single raw frame with properly adjusted illumination levels? Up to this point in time, I can still not give a definite answer, because I simply did not have the time to make some additional experiments. Anyway, what I know so far:
-
A well-exposed single raw camera capture with 12 or 14 bit resolution can capture basically the full dynamical range of a color-reversal film. It is much faster than taking the 5 or so needed, different exposures for exposure fusion. If you go for exposure fusion and take 5 exposures by varying quickly the illumination source, you can work much faster than if you keep the illumination constant and vary the exposure time of your camera. Still, exposure fusion is much slower than a single raw capture. I am currenty able to scan 1 minute of 18fps-film in about 34 minutes.
-
12 to 14 bit is barely enough to capture the full dynamical range of color-reversal film. Slight variations in the exposure of the original film might push your working envelope in such a way that either the shadows are being just pure black or the highlights being burned-out. Whether this is noticable in the final scan remains to be seen - after all, your end product will normally be displayed anyway on a limited 8-bit display.
-
With illumination-optimized raw captures, your original material will be off quite a bit from what you want to achieve for final presentation - thus, extensive (manual) post processing is needed. With exposure fusion, color corrections needed in post production are less severe. Exposure fusion has the tendency of averaging out certain variation - so post production work is reduced. You can see some examples of this “averaging out” of original film variation in the examples I posted here.
-
Exposure Fusion is a computationally expensive algorithm, and it scales badly with scan resolution. I never tried 4K, but with 2012 x 1512 px, I need with my not-so-bad desktop 2.3 seconds to combine a single frame into the output frame. This adds substantially to the overall time you have to invest for scanning a given amount of film. Processing of 1 min of 18 fps film will take about 41 min. That is longer than the actual scanning time! Since this is an automatic process, I usually let the stuff run overnight. In total, my system currently needs at a resolution of 2012 x 1512 px for 1 min of 18fps film a processing time to arrive at the exposure fused base material of 1 hour and 15 mins. Nothing for the impatient!
-
Because a lot of computations are involved, the output image of the exposure fusion algorithm is a little bit less sharp than the image one would obtain from a single raw capture. That can however be corrected by an appropriate sharping step during processing. Furthermore, exposure fusion has the tendency to enhance sligthly local contrast features - this improves overall image perception a little bit. You can see this if you look closely at the original size images I posted above and compare.
-
Exposure fusion does not benefit at all from a finely tuned illumination. That is caused by the huge range of EVs covered. Given, each EV exposure is only output (as a jpg) with 8 bits (per pixel, per color channel), but the number of digits available in the 5 different exposures covers a wider dynamic range, probably equivalent to a raw capture with more than 18 bits (I need to calculate that in more detail - this is just a back-of-the-envelope estimate). You can see this in the examples given here as well. The second and third example given there would have needed very different illumination settings if done with a single raw capture (the second example is very bright, the third example very dark). Exposure fusion just picks the right illumination level automatically. As a matter of fact, the illumination/exposure levels in my approach are set to the maximum intensity of the light source, not at all to what specific film is actually scanned. The values stay fixed whatever I scan.