GH4 : Adam Wilts's review

Very enlightening - thanks for the link, Bern. We know that downconverted 4K looks great, usually, but Adam's quantification of GH3/4's FHD real resolution of 675-700 TVL/ph before aliasing sets in vs. downconverted GH4 4K reaching a full 1080 (all at default sharpening level "0"), giving a potential of over 50% gain in real definition. ETC mode, at 1:1 pixel ratio, goes to 800, and C4K/UHD modes both double this as expected.
 
His reviews are always such a good read. So, 10 stops dynamic range... with a slight chance at pulling out a touch more. Seems in line with what I thought I'd been seeing. Probably good enough for me.
 
For dynamic range, I'd go by DxOMark's numbers or a Xyla chart. The Stouffer wedges are not very accurate, especially as you get down to the darkest blacks.

Imatest does list the Stouffer T4110 as usable for DR testing (with 13.3 stops of range). However, looking at the specs for the T4110 (http://www.stouffer.net/T4110spec.htm) at around 3.05 Dmax (or at the 10 stop mark) the tolerance levels equal the transmission levels (which I'm interpreting to mean that from there on out the numbers cannot be guaranteed to be accurate - which is also evidenced by the fact that "n/a" is listed for the actual % of light transmission for each step after that.) On top of that, for Adam's test, at least the first two steps (2/3rds of a stop) were clipped, which means that the 3.05 step would be reached much sooner.

Someone tested an F55 using Stouffer wedges over at BMCUser and only came up with 10 stops of range, which can't be right: http://www.bmcuser.com/showthread.php?8355-BMPC-4k-Dynamic-Range-Chart/page3 .

However, the most conclusive evidence I've found that shows how inaccurate the Stouffer wedges can be is the fact that DPReview still uses them in every single review, yet not one camera has surpassed the range of the chart yet. Even the Nikon D800 with Active D Lighting turned up to Extra High only shows 11 and 2/3rds stops: http://www.dpreview.com/reviews/nikon-d800-d800e/19.

So, I'm still sticking with 12.8 stops. :)
 
Joshua Cadmium said:
For dynamic range, I'd go by DxOMark's numbers or a Xyla chart. The Stouffer wedges are not very accurate, especially as you get down to the darkest blacks.
Click to expand...

DxOMark is actually fairly accurate for many things, I've found, so I'll take their word on DR, as I don't have a way to properly test DR in an absolute objective manner. However doesn't DxOMark test raw image data? Since video is not raw here, it's going to be compressed data with the tone curve baked in already. I doubt it will do the full 12.8 stops in video.
 
I thought DxO etc was DR in photo RAW mode.......not video?

Either way, I try not to get too hung up clinical tests, my eyeball tells me the GH4 looks pretty decent for DR overall. Definitely better than my GH2. For the money it's hard to complain......although I'll still find something to nitpick over.
 
Great review. Thanks for sharing it. It's nice to finally see some resolution charts and dynamic range tests. And it was nice of the author to reference my moiré test video in his conclusion.
 
yslee said:
DxOMark is actually fairly accurate for many things, I've found, so I'll take their word on DR, as I don't have a way to properly test DR in an absolute objective manner. However doesn't DxOMark test raw image data? Since video is not raw here, it's going to be compressed data with the tone curve baked in already. I doubt it will do the full 12.8 stops in video.
Click to expand...

DxOMark does use raw image data, but the white clipping point (full well capacity) is going to be the same in video, and because the tone curve and master pedestal can move black information higher, information just above the noise floor (read noise) isn't going to mush out and be unusable. (And dynamic range is just full well capacity divided by read noise.) When the tone settings make the shadows noisy, basically all the DR that is there in the raw data is being made available to be recorded by the codec.

Also, video actual has the potential for higher dynamic range than stills. This is because fluctuating noise, like shot noise and photon noise (noise coming from the randomness of electrical current, and noise coming from photons not hitting the sensor at the same time) can be averaged out with temporal noise reduction, like Neat Video. You can't do that with stills, unless you take multiple stills and average them. The net effect is additional dynamic range in the shadows, because you are lowering the noise floor.

Also, the multiple frames of video will average out in your brain, as well. Think about how grainy a single frame of 35mm film can look, but how less grainy it appears in motion. Since grain is random, just like a lot of sensor noise, you see the image presented despite the grain.

So, I'm sticking with 12.8. :)
 
Well, that's perceptual, not absolute, surely? I get what you're trying to say though.

I'm saying this because mostly I know when a camera has crappy DR: Post-processing (grading) its raw images usually shows up the issues when I start messing with the curves and colour information.

It's also why grading compressed footage feels a bit like going back to 2001 for me: JPEG post-processing feels so limiting.
 
Joshua Cadmium said:
DxOMark does use raw image data, but the white clipping point (full well capacity) is going to be the same in video....When the tone settings make the shadows noisy, basically all the DR that is there in the raw data is being made available to be recorded by the codec.

So, I'm sticking with 12.8. :)
Click to expand...

But isn't it the case that RAW images are taken directly from the camera's 14-bit A/D converter, while the video gets encoded into 8-bit? Doesn't this have an impact on dynamic range? I am not an engineer, but the common wisdom about the video modes of DSLRs is that they do not have the dynamic range of RAW stills. I can confirm this with all the GH cameras I own. A frame of video will always clip at a lower setting than a RAW still of the same subject at the same exposure settings.
I believe the 8-bit, 4:2:0 h.264 codec hampers the camera's potential dynamic range. That's why I'm hoping to see a slight improvement when the Atomos Shogun arrives, and we can get 10-bit out of the camera.
 
There is no way that you can say that because the cameras dynamic range is XX stops with RAW stills, that it is the same with compressed video. I still cant find a definitive answer on the GH4's dynamic range in video mode.
 
dvbrother said:
But isn't it the case that RAW images are taken directly from the camera's 14-bit A/D converter, while the video gets encoded into 8-bit? Doesn't this have an impact on dynamic range?
Click to expand...

In terms of detail, color, and tonality, raw is always going to be better.

However, 8 bit doesn't have less dynamic range (as we're talking about it) than raw, it has less tonality, which means that the whole image has less tonal detail, including the shadows and highlights. Look at the images at the bottom of this page: http://www.phototechnique.com/how-to/practical-guide-to-dynamic-range/ . The 1 bit image looks completely blown out, but that doesn't necessarily mean it has less dynamic range (as we're talking about it), just that it doesn't have enough available tones to fully describe the detail.

In order to define dynamic range a bit better for those who aren't aware, there are two types of dynamic range at play here. Sensor dynamic range, which is what we're really concerned about and which doesn't really change regardless of how it is recorded, and tonal dynamic range. Sensor DR is defined by full well capacity, or clipping point, divided by the noise floor, or read noise. Tonal dynamic range is defined by the amount of possible tones in an image.

So, 8 bits of data can only technically record 8 tonal stops of dynamic range - 2^8=256. And you would get 8 stops of sensor dynamic range if the 14 bit to 8 bit downconverter just chose 8 out of the 14 bits worth of information somewhere in the middle of the range (just taking a slice of the data.) However, the downcoverter does use all 14 bits worth of information and shapes it into a data range that will fit within 8 bits (shrinking the pie instead of just taking a slice of the pie.)

Therefore, the very brightest bits and very darkest bits of sensor dynamic range are preserved (which is what were concerned about) just with less tonal steps.

Now, less tonal dynamic range does appear to be less sensor dynamic range. If you were looking at a bright cloud, detail just below clipping in 14 bit land can only choose one of two possible options in 8 bit land - 254 or 255. So, you could say that there is less sensor dynamic range in 8 bit, since a bright highlight that wouldn't clip with 14 bits is now clipping with 8 bits (anything over 244.5). But, for highlight and shadow that's only 2 clipping points in 256 possible tones, or 128th less dynamic range than raw. So 12.8 * 127/128 = 12.7 stops of DR, which is close enough to 12.8 to be negligible. (However, we're really splitting hairs here. By using the same math you would say that a 1 bit image has zero dynamic range.)

The great thing about this camera is the options available to use 8 bits worth of data as efficiently as possible in order to maximize those 256 steps of tonal range. Those tonal settings are likely coming directly from raw data, so you're shaping those 14 bits before they're being downconverted to 8 bits.

So, in 8 bits, there is less tonal dynamic range, but sensor dynamic range is the same, or as close as it's going to get in 8 bits, which we have a lot of control over, which means that 12.8 stops of DR is still a valid number. :)
 
Last edited:
Joshua Cadmium said:
8 bits of data can only technically record 8 tonal stops of dynamic range - 2^8=256. And you would get 8 stops of sensor dynamic range if the 14 bit to 8 bit downconverter just chose 8 out of the 14 bits worth of information somewhere in the middle of the range (just taking a slice of the data.) However, the downcoverter does use all 14 bits worth of information and shapes it into a data range that will fit within 8 bits (shrinking the pie instead of just taking a slice of the pie.)

Therefore, the very brightest bits and very darkest bits of sensor dynamic range are preserved (which is what were concerned about) just with less tonal steps.
Click to expand...
Good summary, just a couple of corrections:

According to DxOMark, the GH4 has a 12-bit sensor, not 14-bit. This is how earlier GH models worked as well.

When that 12-bit dynamic range is compressed into an 8-bit encoding, there are a lot more tonal steps available at the bright end of the scale than at the dark end. What primarily reduces highlight tonality depth is the type of gamma curve used in the selected picture profile. Panasonic documented how their Cine-D, Cine-V, and B.Press gamma curves work here:

https://eww.pass.panasonic.co.jp/pro-av/support/content/guide/EN/AG-HMC150handbook_e.pdf

While that link specifically refers to the HMC150, my impression is that Panasonic's gamma curves are standardized among camera models. If you want the TL;DR on highlights, Cine-D will give you more highlight detail than Cine-V or B.Press.
 
Joshua Cadmium said:
In terms of detail, color, and tonality, raw is always going to be better.

However, 8 bit doesn't have less dynamic range (as we're talking about it) than raw, it has less tonality, which means that the whole image has less tonal detail, including the shadows and highlights. Look at the images at the bottom of this page: http://www.phototechnique.com/how-to/practical-guide-to-dynamic-range/ . The 1 bit image looks completely blown out, but that doesn't necessarily mean it has less dynamic range (as we're talking about it), just that it doesn't have enough available tones to fully describe the detail.

In order to define dynamic range a bit better for those who aren't aware, there are two types of dynamic range at play here. Sensor dynamic range, which is what we're really concerned about and which doesn't really change regardless of how it is recorded, and tonal dynamic range. Sensor DR is defined by full well capacity, or clipping point, divided by the noise floor, or read noise. Tonal dynamic range is defined by the amount of possible tones in an image.

So, 8 bits of data can only technically record 8 tonal stops of dynamic range - 2^8=256. And you would get 8 stops of sensor dynamic range if the 14 bit to 8 bit downconverter just chose 8 out of the 14 bits worth of information somewhere in the middle of the range (just taking a slice of the data.) However, the downcoverter does use all 14 bits worth of information and shapes it into a data range that will fit within 8 bits (shrinking the pie instead of just taking a slice of the pie.)

Therefore, the very brightest bits and very darkest bits of sensor dynamic range are preserved (which is what were concerned about) just with less tonal steps.

Now, less tonal dynamic range does appear to be less sensor dynamic range. If you were looking at a bright cloud, detail just below clipping in 14 bit land can only choose one of two possible options in 8 bit land - 254 or 255. So, you could say that there is less sensor dynamic range in 8 bit, since a bright highlight that wouldn't clip with 14 bits is now clipping with 8 bits (anything over 244.5). But, for highlight and shadow that's only 2 clipping points in 256 possible tones, or 128th less dynamic range than raw. So 12.8 * 127/128 = 12.7 stops of DR, which is close enough to 12.8 to be negligible. (However, we're really splitting hairs here. By using the same math you would say that a 1 bit image has zero dynamic range.)

The great thing about this camera is the options available to use 8 bits worth of data as efficiently as possible in order to maximize those 256 steps of tonal range. Those tonal settings are likely coming directly from raw data, so you're shaping those 14 bits before they're being downconverted to 8 bits.

So, in 8 bits, there is less tonal dynamic range, but sensor dynamic range is the same, or as close as it's going to get in 8 bits, which we have a lot of control over, which means that 12.8 stops of DR is still a valid number. :)
Click to expand...


Hmm, not sure I agree with you.

Basically you are saying that the clipping point of white and black remains the same wether taking RAW stills or compressed video but that there is a much finer graduation (smaller steps) in the RAW image rather than more dynamic range?

In practice, shooting the same scene, I can pull a lot of detail out of the shadows and highlights in the RAW image that just isn't there on the MOV files. While visually the dynamic range might be similar (if not the same), once you start manipulating the image you find that the RAW image can see much further into the shadows and highlights.
 
vtwiz said:
Hmm, not sure I agree with you.

Basically you are saying that the clipping point of white and black remains the same wether taking RAW stills or compressed video but that there is a much finer graduation (smaller steps) in the RAW image rather than more dynamic range?

In practice, shooting the same scene, I can pull a lot of detail out of the shadows and highlights in the RAW image that just isn't there on the MOV files. While visually the dynamic range might be similar (if not the same), once you start manipulating the image you find that the RAW image can see much further into the shadows and highlights.
Click to expand...

Raw definitely has more data in the shadows and highlights. And that is indeed because there are more gradation steps to define that data.

8 bit has at most 256 distinct steps per color (2^8), while 12 bit can have up to 4,096 distinct steps per color, and 14 bit will max out at 16,384 distinct steps per color.

If your delivery only needs to be in 8 bit, but you have 12 bit raw data, those 4,096 tones (including much more in the shadows and highlights) can be stretched out without causing banding. That is what you're really seeing with a native 8 bit file once you start trying to pull up shadows and pull down highlights. It's just banding, which looks a lot like clipping, and which can also be exacerbated by the codec trying to efficiently throw away data.

The point I was trying to make was not that raw has more data, it definitely does, or that 8 bit compressed video has as much gradeability as raw (it doesn't - although I think the 4k 100Mbps of the GH4 has quite a bit of leeway). I was just trying to point out that raw stills and video has the same sensor dynamic range (12.8 stops.) You could say that it feels like it has less dynamic range (and I have heard people use the term "perceived dynamic range" when describing higher bit depth video) but that doesn't mean that you could say that the actual sensor dynamic range numbers have changed.

Otherwise, you could say that all 8 bit video (0-255) has the same dynamic range. Tone-wise that would be correct, there are only 255 steps, so all 8-bit video only has 8 stops of tonal dynamic range, including the GH4, the T2i, the C300, and the A7s.

What we're really concerned about, though, is how the light in the scene we're shooting is going to fit into those 8 bits through the eyes of the sensor.

With the GH4, we know that 12.8 stops of light can be seen by the sensor. So, if your scene is equal or less than that, you're fine, otherwise you need to decide how you're going to expose to record the portion of dynamic range you want. Or you can light the scene a bit better, if possible, to fit into that range.

Also, don't forget, while raw has more detail in the shadows and highlights, with the tone curves on the GH4 you can bring out that detail before it's compressed to 8 bit video. This is why some of the combinations of tone settings have been noisy - you're seeing additional shadow information, potentially down to the noise floor.

A lot of people over the years have said that raw has more dynamic range - and it does - but only tonally. For instance, check out this test on the 550D (T2i): http://sareesh.com/2011/07/dynamic-range-comparison-of-raw-vs-video-mode-on-the-canon-550d/ . Sareesh (who I really respect - I love wolfcrow) comes to the conclusion that the raw photo has more dynamic range. But look at the two pictures initially - white is clipping at the same point and shadow detail is exactly the same. When he brings out detail two stops above and below, both have the same details - they are just better described in the raw file because there are more tones to do so.

So, raw does have more detail you can pull out, but it's clustered about the white clipping point and black noise floor point - which are the same in video mode as well.

What I'm saying is just that if you're going to pick a number to describe the sensor dynamic range of the camera, it's going to be the same regardless of how the image is recorded (raw, prores, tiff, jpeg, avchd, etc.) Some of those formats are going to give you more information and thus more leeway when grading, but they all are coming from a sensor with the same dynamic range.
 
You must log in or register to reply here.
Back
Top