P4P, can we squeeze 14 stops of dynamic range from a raw file?

[Tomas Wangen]

Yes, yes we can.

Please note that this is a ad-hoc experiment I did to just see how much it was possible to get out of a ISO 100 capture.

Measurements are done with reflected light measurements from a Gossen Digisky which will add a slight error as compared to what the sensor sees, let us assume this is up to 1 stop (guess from my side).

Still, the results are rather impressive.

On the sample below, I overexposed with 1 stop according to the dji go 4 histogram, but were able to resolve all the whites (using -100 for white in lightroom). Heavy noise-reduction applied (to get usable shadows): 32 luminance, 70 color.

The (b)right side is illuminated by two 1500 lumen lume-cubes while the left has room-lighting, so there is a colorshift (not corrected, only global adjustments used).

For reference, default render from lightroom (except colortemperature) is below.

Images are a 100% crop (all pixels present).

For further reference, here is the same capture as jpeg/dlog -1/0/0. We have lost 1 stop of highlights and about 1/2 or 2/3 stops of shadows as compared to raw, so dynamic range here is roughly 12.5 stops (which will be the same with video / dlog).

Worth noting is that dlog has preserved the shadows very well.

 

[Raptorman0909]

Rest assured, if the camera were capable of delivering 14 stops of DR that DJI would be saying that -- they only claim 12 stops. The luminance patches I would guess are spaced at 1/3 stops and I count 11 steps or a difference of 3 1/3 stops not 5 stops. I guess they could be spaced at 1/2 stops and that would equate to 5 stops total, but 1/2 stops is non standard.


Brian

 

[Roamer105]

You might get 14 stops or possibly even more with HDR capture.

 

[Tomas Wangen]

Rest assured, if the camera were capable of delivering 14 stops of DR that DJI would be saying that -- they only claim 12 stops. The luminance patches I would guess are spaced at 1/3 stops and I count 11 steps or a difference of 3 1/3 stops not 5 stops. I guess they could be spaced at 1/2 stops and that would equate to 5 stops total, but 1/2 stops is non standard.


Brian

The measurements are ofcourse done directly on the prints during capture except much closer than the sensor. I have not counted stops, these are reflected light measurements in the indicated spots. Also note that I overexposed with one stop (this was lost in the reference jpeg / dlog) but was able to get this back by correcting whites-clipping.

Also note that the noise is extreme in the shadows (after lifting everything) so this is not something you would do other that in a experiment such as this (I assume).

 

[Andrei V]

Yes, yes we can.

Please note that this is a ad-hoc experiment I did to just see how much it was possible to get out of a ISO 100 capture.

Measurements are done with reflected light measurements from a Gossen Digisky which will add a slight error as compared to what the sensor sees, let us assume this is up to 1 stop (guess from my side).

Still, the results are rather impressive.

On the sample below, I overexposed with 1 stop according to the dji go 4 histogram, but were able to resolve all the whites (using -100 for white in lightroom). Heavy noise-reduction applied (to get usable shadows): 32 luminance, 70 color.

The (b)right side is illuminated by two 1500 lumen lume-cubes while the left has room-lighting, so there is a colorshift (not corrected, only global adjustments used).

For reference, default render from lightroom (except colortemperature) is below.

Images are a 100% crop (all pixels present).

View attachment 73522

View attachment 73508

For further reference, here is the same capture as jpeg/dlog -1/0/0. We have lost 1 stop of highlights and about 1/2 or 2/3 stops of shadows as compared to raw, so dynamic range here is roughly 12.5 stops (which will be the same with video / dlog).

Worth noting is that dlog has preserved the shadows very well.

View attachment 73516

Tomas, is this with the latest firmware? Did it improve DLOG finally?

Also, could you perhaps repeat the experiment and use a frame grab from video rather than the JPEG for comparioson with the raw do see if any of that dynamic range is lost in the video encoding as apposed to JPEG encoding?

(of course, comparing NONE/DCINELIKE/TRUECOLOR under the same conditions would also be neat, if you have the time)

 

[Tomas Wangen]

Tomas, is this with the latest firmware? Did it improve DLOG finally?

Also, could you perhaps repeat the experiment and use a frame grab from video rather than the JPEG for comparioson with the raw do see if any of that dynamic range is lost in the video encoding as apposed to JPEG encoding?

(of course, comparing NONE/DCINELIKE/TRUECOLOR under the same conditions would also be neat, if you have the time)

Hi, the last image is dlog which i measured to about 12.5 stops. Dlog now works (i have some results in the "new firmware" thread.

The other color profiles probably have the same dynamic range (black to white clipping points) but will suffer in the shadows compared to dlog.

 

[Andrei V]

Tomas, is the last image DLOG from an in-camera JPEG, or from an in-camera h.264/h.265 recording? In camera JPEGs and video recordings may not behave in the same way.

 

[Tomas Wangen]

This test was for raw (so same capture). I would be very surprised if they have used different curves for video but this is of course easy to verify.

 

[Andrei V]

I suspect the hardware path for JPEG encoding and h.264/h.265 encoding is entirely different (i.e. different chips or different hardware logic blocks are used), so it's quite possible that the gamma curve applied and recorded is different enough to lead to different dynamic range results. While both the JPEG and h.265/h.265 files are 8-bit files, the different compression schemes and the degree of compression could also impact on the actual noise and retained dynamic range.

I'd test it myself, but can't reproduced the control lighting environment and step chart you have easily

 

[Tomas Wangen]

I suspect the hardware path for JPEG encoding and h.264/h.265 encoding is entirely different (i.e. different chips or different hardware logic blocks are used), so it's quite possible that the gamma curve applied and recorded is different enough to lead to different dynamic range results. While both the JPEG and h.265/h.265 files are 8-bit files, the different compression schemes and the degree of compression could also impact on the actual noise and retained dynamic range.

I'd test it myself, but can't reproduced the control lighting environment and step chart you have easily

To verify: take a 16/9 picture indoors using controlled light (for instance a lamp) using dlog. Take a video same scene. Compare the resulting histograms. Only requirement is manual exposure to ensure same settings.

 

[Tomas Wangen]

Now I could not help get a bit curious about this myself, so I will do the test later today.

 

[Andrei V]

To verify: take a 16/9 picture indoors using controlled light (for instance a lamp) using dlog. Take a video same scene. Compare the resulting histograms. Only requirement is manual exposure to ensure same settings.

Hmm, yeah, I suppose that should work. I'd probably pull a frame-grab from the video as a TIFF file or something and compare it to the in-camera JPEG in Lightroom or similar.

 

[Tomas Wangen]

Hmm, yeah, I suppose that should work. I'd probably pull a frame-grab from the video as a TIFF file or something and compare it to the in-camera JPEG in Lightroom or similar.

I plan to import both in davici so I can be absolutely sure colorspace is not adjusted during any conversion.

 

[Tomas Wangen]

So, I finally did the controlled test between jpeg dlog and mp4 dlog. I expected to find exactly the same but had to double and triplecheck as i did not.

What we see is a slight shift in color/tone between them, but it is almost imperceptible and probably not of any effect. (checked in both photoshop and resolve with exactly same result).

I expected initially to find about 12.5 stops of usable dynamic range, and based on these results i will say 12 is a fair estimate, which corresponds with djis promises and inspires confidence.

From a technical standpoint, these are two separate captures for dark and light, with same exposure and different lighting. I tried to match the darkest vs lightest patch.

I am happy with the results and will say dji has delivered as promised.

(note: if it seems there is more to get than the 12.2 swatch, this is were the blue channel clips)

 

[ianwood]

For such a small and cost effective drone, 12 makes me quite happy.

 

[Roland Schulz]

Hi Tomas,
nice work here, but I expect the P4 Pro to be around 12-13 stops according to what DXOmark measured for several Sony RX10/100 cameras that contain a similar sensor (P4 Pro has Sony IMX183 Sensor).
On the other hand it´s nowhere defined which values to count for measurement and where to stop. If a difference is visible I would also count it as a value.
I will also measure DR of several cameras soon building an own DR wedge like the DSC labs XYLA. I already ordered some ND foils from LEE and will verify the value of overlayed foils with a gossen light meter.

 

[Raptorman0909]

I posted another thread on my testing of D-Cinelike and D-Log after the new firmware of a few days ago. It looks like DJI has changed the way in which the histogram and video feed is displayed when using D-Log and that it is now easier to read and set an exposure, but I am not pleased with the quality of video D-Log is providing. I'm getting severe banding in the sky while using D-Log and I do not see the banding with D-Cinelike. Both D-Cinelike and D-Log have the flickering effect though it is more noticeable with D-Log, I think because you need to use more contrast stretching with D-Log. I will try to put together a video to demonstrate this tomorrow and may provide the rendered video and one of the raw video files on my Google Drive so you can see it before YouTube washes it with there process.

So, while the early reports had me hoping the D-Log was fixed I can tell you that it is not by my eyes.

Testing the new firmware -- D-Log and D-Cinelike


Brian

 

[Tomas Wangen]

This particular thread and the analysis and results is about the maximum retrievable dynamic range of one still. I have not had the weather to do a complete video analysis yet.

However I must say that based on the new dlog profile itself, there is nothing to suggest new banding problems so I am surprised with your findings.

Looking forward to your examples.

 

[Tomas Wangen]

Hi Tomas,
nice work here, but I expect the P4 Pro to be around 12-13 stops according to what DXOmark measured for several Sony RX10/100 cameras that contain a similar sensor (P4 Pro has Sony IMX183 Sensor).
On the other hand it´s nowhere defined which values to count for measurement and where to stop. If a difference is visible I would also count it as a value.
I will also measure DR of several cameras soon building an own DR wedge like the DSC labs XYLA. I already ordered some ND foils from LEE and will verify the value of overlayed foils with a gossen light meter.

Looking forward to your findings!

I already mentioned this, but to get to 14 stops I had to overexpose and recover whites which is only possible with raw.

 

[Andrei V]

So, I finally did the controlled test between jpeg dlog and mp4 dlog. I expected to find exactly the same but had to double and triplecheck as i did not.

View attachment 73747

What we see is a slight shift in color/tone between them, but it is almost imperceptible and probably not of any effect. (checked in both photoshop and resolve with exactly same result).

I expected initially to find about 12.5 stops of usable dynamic range, and based on these results i will say 12 is a fair estimate, which corresponds with djis promises and inspires confidence.

From a technical standpoint, these are two separate captures for dark and light, with same exposure and different lighting. I tried to match the darkest vs lightest patch.

I am happy with the results and will say dji has delivered as promised.

(note: if it seems there is more to get than the 12.2 swatch, this is were the blue channel clips)

Thanks for updating this, Tomas. I figured the mpeg might be retaining slightly less dynamic range then the jpeg, and I guess it is, though the difference is small enough to be largely inconsequential. 12 stops of dynamic range is definitely nothing to complain about, and would be hard to better -- the m4/3 sensors on the inspire x5 line probably only get you about 1 stop more if shooting in raw/prores.

It's nice to see dlog is now a useful tool again, even if it has downsides (like color compression and subsequent banding when grading) that might advocate against it's use in every circumstance.