April 21, 2017 at 2:51PM


How a 10 bit file can store all of the detail from a 14 bit sensor

I have a post here https://plus.google.com/108780727762043742679/posts/1mE8dNbeMbA
where I explain a bit about how 14 bits of linear image data can stored in a 10 bit file without losing real image details.

It is to explain a bit why camera's with 14 stops of range such as the new Panasonic can safely store their data in just a 10 bit file.

If you all could take the time to see if it makes sense to you I'd appreciate it.


The simple answer is: you cannot store 14 bits of linear data into 10 bits without some loss. However, the eye is not a 14-bit calculator with an answer key ready to mark wrong every pixel that is not perfectly represented in the 10-bit world. Our eyes use a log-based valuation system, and linear encoding greatly over-samples highlights while substantially under-sampling shadows. Converting linear data to log data throws away details in the highlights that don't matter while faithfully representing all the intrinsic noise and uncertainty that exist in the shadows (where more than half of the 10 bits of log data representing different stops are all the same--zeroes).

Now, where all that highlight stuff does start to matter is when you want to wildly change your ISO interpretation of the video. With the RED camera, you can re-interpret RAW data. If you expose for ISO 3200 and "develop" at ISO 100, the stuff that was all supposed to be clipped at white is now 5 stops lower down the exposure ladder. And you want to be able to see the distinction between 100%, 90%, 80%, 70%, and 60% white. That would be impossible if the DR is first squeezed into 10 bits.

There very much is a limit to how much one can do with a reduced-bit encoding, both in terms of absolutes as well as the approximations between the stops. If your exposure is bang-on for the 14 bits you want to record, your eye will forgive having only 0.7 bits of accuracy per stop. If you want the flexibility to play with exposure, then you really want at least 1 bit per stop.

April 22, 2017 at 8:05AM


Yeah more bits mostly better, but your example seems to be deliberately processing an image incorrectly that has been recorded properly.

If you take a red camera and set the iso at 3200. it records the image the same as if you set it to iso 100. All you are doing is picking where the camera changes the camera tap out to. The recorded files only change the metadata. It sounds like you are deliberately picking a setting that will waste 5 stops worth of data if written to file that way, without any reason.

If you are exposing an image to preserve highlight data you would expose the to preserve it, overrate your camera. Then the resulting files is written out at the protection iso(3200). That preserves all of the data from the sensor, the final exposure gets changed in the DI platform.

The iso setting on the red software is just a post process on the recorded data, If you set it incorrectly you can get bad results.

Also why would 1/2 the bits in a 10 bit file be 0's? That isn't true at all for the redLogFilm format. RedlogFilm actually shows how how extremely well 14 stops fit into 10 bits.

You would never want to grade or post process in 10 bits but it is a perfectly reasonable format for a 14 bit camera to record in. I.e. the gh5.

April 25, 2017 at 3:25AM


Your Comment