HDR+0

[Robert Hofmeyr]

Thanks John.

 

[Lee Saxon]

Robert, I think shooting 50p instead of 25p is what you're looking for.

 

[Justin O'Neill]

Very interesting Les and Blair. Looking at my HDRx footage it appears they are consecutive as well. The beginning of the A track definitely starts after the end of the X track.

Just a thought, but who says the simultaneous exposures START at the same time? Perhaps they end at the same time...so the X track is using an exposure differential that puts its measurements in the higher total cumulative values instead of starting in the noise floor and taking a quicker 'snap' that is close to the noise threshold of the sensor?

You can tell by looking at where objects are in the frame that the X track definitely happens before the A track.

 

[Matt Gerard]

Dang, I could have confirmed this, yes, the X track gets recorded first, then the a track. The RED 220 course at fxphd run by Mike Seymour really put a lot of technical stuff in layman's terms for me. If you look in the HDRx settings in RCX, there is the Simple Blend and Magic Motion options. The 2nd option is the magic sauce they are using to make the temporal retiming a bit more natural. To my eyes it is very subtle, but its there. Think of a shot of a person standing in front of the camera waving thier arms. In standard motion blurred video, the arms are totally blurred, no sharp edges except for the leading edge depending on which way the arm is moving. Now, shoot this with HDRX, and the sharper non-motion-blurred x track. Simple blend these 2 together and you get an overlapping effect that doesn't look natural, it looks as if one is laid over the top of the other, so you see the entire sharp hand moving (xtrack), along with the motion blurred track (a track). The magic Motion option is supposed to (apparently) retime these a bit and blend them in a little different way to get a more natural blend and retiming. Like I said its hard for my eye to see it, but I can see it when it is compared side by side. I'd love to post some clips from the training course, but I dont want to get in trouble for it! FXPHD did a great job of explaining it.

 

[Les Dittert]

I hope I put the 'simultaneous exposure' of A and X tracks to bed forever with the example images.
I am a strong believer in scientific methods to reveal information and facts.
Anyone can repeat the test, with a record player spinning an object, cars going down a fast road, etc.

Another thing: I was led to believe that there was perhaps some special processing in RCX to realign the A and X tracks to avoid artifacts. Simple case: doing a pan on a high contrast edge of a window frame. I have not seen evidence of this yet. I did the align in post myself. My fix only works on simple motion, BTW. I didn't see any 'magic' in the RCX processing.

I'm not complaining at all. HDRX can be used if you are careful. No other movie camera has such a feature, I commend Red for adding it in there.

 

[Terry VerHaar]

So - it does feel a little like RED has been somewhat misleading on this front. Of course, it is always possible that I misinterpreted but the Ops Guide does say "simultaneous" and, from memory (which of course, is always suspect) I seem to remember the RED folks saying the HDRx exposures were definitely NOT shot sequentially but AT THE SAME TIME.

I just went to try to shoot HDRx with 360 degree shutter and, indeed, it appears it can't be done.

At this point, from all I can discern, I was certainly flat out wrong, despite all the evidence that supported my (supposed) knowledge.

It really would be nice to hear something definitive from RED at this point. I do sort of feel like the value of HDRx is somewhat diminished under these conditions. :-(

 

[Blair S. Paulsen]

Based on observations noted in this thread it seems the X track exposure is first, then the A track exposure. I still contend that since the time interval between them is less than 10ms (AFAIK) then for a lot of scenes its going to look the same as if there was a way to read out the registers for the X track while still accumulating greater voltages for the A track readout (which I have been told is not technically feasible at this time, but I lack direct knowledge).

In any case, when RED first enabled HDRx I had a buddy of mine who is a mondo rock drummer set up his kit in open sunlight and whale on it. His hands are moving fast, the drumsticks even faster, the elbows a little slower and the shoulders a bit slower than that (some exceptions, but lets not digress). I was looking to see what would happen if the motion blur on the A track was extreme enough that the object, in this case the drumstick, would leave only a semi-transparent wash while the X track would actually show the drumstick and yield a type of judder effect. I also wanted to see how the different speeds at which various parts of the drummer and his sticks moved would change the character of the "blend" depending on levels of motion blur.

My goal was to understand the practical limitations of HDRx so I could provide meaningful advice on what to expect when using it in various situations. As Jim has mentioned in past posts you have to start with breaking down how the eye/brain combination actually perceives and processes visual stimuli, specifically when observing fast moving objects. He contends, and I concur, that both "frozen" images of the fast moving object and blurred versions are being "input" into your visual cortex and "resolved" into a hybrid perception of the object in your brain.

Jim, if I have mis-characterized your position on this my apologies, I am just trying to inject the idea that there are several interactive factors at play.

Cheers - #19

 

[Robert Hofmeyr]

Okay, we understand how it works now. Technically, I assume enabling HDRx+0 would be trivial. Would it work as I assumed in the OP? Would it be useful? It would add a 1stop "digital ND" which would be marginally useful when stuck without NDs or when using a lot of NDs and needing to limit IR pollution. Additionally, either of the tracks could be used on their own for sharper stills or motion tracking. The sharper track would have the same exposure as the combined track (which would be great) but of course your stills would only have double the apparent shutter speed of the combined track, so the usefulness would be limited.

 

[Robert Hofmeyr]

Lee, I'm not sure how 50p would help. If I want 25p 180 degrees and shoot 50p 360 deg, all I could do is discard every other frame. If I shoot 50p 180 deg, the frames could not be joined to make 25p 180deg.

 

[Blair S. Paulsen]

Changing the shutter angle (or speed) can be an alternative to using ND filters - of course the blur characteristic will change, which often results in very different looking motion. This will be the case whether you use HDRx to thin the shutter or simply change the shutter directly - no free lunch.

FWIW if a full matte box and filter set is not available might I suggest a screw on variable ND filter. Yes, even my Heliopan fader ND introduces some optical degradation. That said, its small, very low cost by comparison, fits in your pocket and with a couple of adapter rings can fit most lenses. Just sayin'

Cheers - #19

 

[Robert Hofmeyr]

Blair, would 2 consecutive 90 degree frames merged together not look exactly like one 180 degree frame with a 0.3 ND?

I do understand that NDs are by far the best solution for this. I don't think it's a replacement, but it may be a useful option.

 

[Blair S. Paulsen]

Theoretically with a static subject and no camera movement you could slice the pie a number of different ways and get the same result.

In your example I think the exposure would be the same, not .3 ND (1 stop) less, but perhaps I am not grasping exactly what you are proposing.

FWIW, if you have no other way of cutting exposure during acquisition other than thinning the shutter and need a finished product that has more typical motion blur you can always add motion blur in post - of course that's still going to look a bit odd and take significant resources on anything but short clips.

Just to satisfy my curiosity, why are you so pumped up to avoid ND filtration? I've made some assumptions on that score but perhaps there are motivations I don't comprehend...

Cheers - #19

 

[PatrickFaith]

I think this hdr0 is a great idea. would allow two stops into the blacks on the current epics with a firmware change (probably 4k though) ... How i do the math is:

For a baseline, 24 fps at 180 shutter simulated using hdr0 ... would shoot at 48 fps at 360 with no increase in stops
frame_hdr[j++] = frame;
i= i +2; // waist every other frame ...

for a 1 stop increase at 24 fps @ 180 shutter , then you'd shoot 96 fps at 360 shutter [edit - bit less then 1 stop ... averaging this noise isn't optimal ... doesn't give a full stop]

frame_hdr[j++] = (frame+frame[i+1)/2;
i= i +4; // half the frames are wasted now i.e. 96 fps gives you 1 stop into the blacks ...

[ i.e. current epic owners can now get one stop into the blacks this way without any firmware change, just a python change in post processing]

if a firmware change was made so the "wasted" frames were not written to the ssd, then you could get "almost" two stops into the blacks (at 192 fps/96 fps):
frame_hdr[j++] = (frame+frame[i+1] +frame[i+2]+frame[i+3] )/4;
i= i + 4;

_____
frame_hdr and frame would have to be 32 bit floats and "frame[]" needs to be linear and the "+" in the example is a matrix add. Would be a lot of work to get two stops into the blacks but "stops" into blacks are "gold", i think this is what is so cool about the Dragon ... is that it will get these stops for "free".
I'm perty sure i did this right ?

btw ... this oversampling basically denormalizes the bayer pattern (i.e. gives you a full sample of blue and red), also since it averages the edges a 4k has almost a 6k feel.

 

[Robert Hofmeyr]

Blair, I'm not pumped up at all, more curious to know if it would work as I imagine. In fact, for my purposes, I think I'd hardly ever use it as a digital ND - the additional processing and dealing with double the data rate would be too large a sacrifice for one less stop of light. I would use it occasionally if I knew I needed to pull stills from the motion - I'd love to be able to pull a still from motion with the same exposure (but half the motion blur).

Patrick, thanks for the clear explanation. When you say "denormalizes the bayer pattern" are you saying that the average of 2 consecutive 1/100sec exposures would have more resolution than a single 1/50th exposure?

 

[Ian Vertovec]

Blair, would 2 consecutive 90 degree frames merged together not look exactly like one 180 degree frame with a 0.3 ND?

I do understand that NDs are by far the best solution for this. I don't think it's a replacement, but it may be a useful option.

Hey Robert, the short answer is no, they would not look the same. It would probably more closely resemble a double exposure of two sharp frames then one nice motion blurry frame. There are very render intensive motion estimation applications like twixtor or optical-flow that would do a better, but not perfect job.

Another issue is you would need to compensate somehow for doubling your data-rate (via frame rate) by either reducing resolution or adding compression, this would lower your overall quality.

Clever idea though. It might be an interesting way to create a custom look, but much less efficient then using an ND. Kinda like how music videos shoot everything overcranked so they can freely create speed ramps in editorial.

 

[PatrickFaith]

Patrick, thanks for the clear explanation. When you say "denormalizes the bayer pattern" are you saying that the average of 2 consecutive 1/100sec exposures would have more resolution than a single 1/50th exposure?

Like Ian is saying ... there's a bunch of wierd things going on when you oversample ... so two 1/100's is not equal to the information of one 1/50'th ( with super cooled film it is the same btw, but in a weird way with analog to digital converters ... there's a bunch of noise in the blacks ... so what your doing is "shifting" down on the d/a and averaging that shift]. I'm doing this from memory of almost 30 years ago ...

The bayer patten means they under sampled red and blue ... but since you got 4 time the amount of sensor readings, you could re-mean them so you would have 100% sensor ratio from rgb( how i think of it in another weird way is i am dropping 3 "greens" from the color calc but still using them to mean the resolution so the 4k is a "real" 4k [i.e. with the current sensor i think a 5k wavelet gives a a "full" 4k resolution ... RED did some test on this a while ago ... can't remember where they are]).

How i explained it btw ... gives no highlight protection ... which is btw what the current hdrx does, only reduces noise in the blacks and averages the resolution to give a cleaner final.

 

[Robert Hofmeyr]

Patrick, I'm not sure I fully understand all of the above but is it correct to say that 2 merged tracks of random noise have less noise than either of the original tracks?

 

[Robert Hofmeyr]

the short answer is no, they would not look the same. It would probably more closely resemble a double exposure of two sharp frames then one nice motion blurry frame.

Thanks Ian, I was under the impression (I think from a RedUser thread years back that I can't seem to find) that they would be basically the same with the end of the motion blur of the first track meeting the beginning of the motion blur of the second.

Good point re lower quality as a result of the added compression, but, if I understand Patrick, you may get back even more quality/res as a result of each frame being a composite of the 2 tracks.

 

[PatrickFaith]

Good point re lower quality as a result of the added compression, but, if I understand Patrick, you may get back even more quality/res as a result of each frame being a composite of the 2 tracks.

Yep, in the old old days this was A LOT more obvious. I.E. pre-cmos (i.e. ttl or matched transistors using compton effect/etc ... ) when using a 4 bit or 6 bit D to A:

bottom bits could average the sensor sensitivity( i.e. a bit flipping back and forth means it's on the edge of 0 and 1 which was mixed with sensor noise), the top bits where used to "uprez" spatially(i.e. if the top bits are flipping then it's on a spatial edge of black and white or the camera is moving which was viewed as spatial noise), and "stuck" bits (bits that don't change with the lower bit changing more then the higher bits), would give sensor error calibration(kind of like blackshading).

In old days there would be a R/C circuit on a different clock that would do the actual "averaging" of the bits with a separate A/D off of the R/C circuit(i.e. so the averaging didn't require any extra samples to be sent). Was obvious to see this since the data rates were not that bad, you could physically point at the circuit, and you could watch it on a oscilloscope. This was back in the early 70's and 80's ... but the basic physics is still the same, it's so embedded now that all this stuff is hidden(i.e. if your not seeing sensor noise on lows, means somewhere in that chain they are sampling to reduce noise and/or adding sensitivity -i.e. on a non-raw camera most of this stuff is said to be "compressed" out, but this actually decreases information so increases error). Fancier circuits then this, in time domains, used exponential smoothing with impulse response to give even better results(i'm assuming that is what is in these modern sensors, but i have no idea since i haven't looked at it for 30 years).

 

[Robert Hofmeyr]

Thanks Patrick. Do you agree that 2 consecutive 90 degree tracks merged together into a single 180 degree track would look stuttery? Has anyone tested this?