It consist of a stepped gradient in the background. That's overlaid with three overlapping circle figures. The one on the left starts at black and moves to more bright in 3% steps (HSL). The round figure on the right starts at black and moves upward in 5% steps. The circles in the center start at white and move downward in 5% steps. Lying over all this is a matrix of black and gray lines on the left and top. And there is a matrix of many white lines on the right. Let's see what Photomatix does with this on its default settings. This is Photomatix Pro version 3.2.7 for a Windows 7 machine. Here it is with default settings.
When I first saw this I thought it was a quilt. We see the same flaring that we saw in the last post. I found another site where this behavior was identified as 'pinching'. On that site, Outback photo, they said this:
"Pinching artifacts reduced (slight change in the Details Enhancer algorithm to avoid the problem of "pinching" artifacts showing along straight lines against a uniform background."
It looks like Outback posted this in the middle of 2008 when they were reviewing 3.0. If this is improvement, well, ...
Also we see the exaggerated shift in toning emphasis whenever a straight line is crossed. This is what I called 'compartmentalization' in the last post. I thought some more about this and I have some ideas about it. Details Enhancer is a local toning algorithm. Speaking in general such algorithms use a radius parameter to define how large an area to process; they then move and, using the same radius, start again and process the new area. The drawback to this algorithm is that sometimes the radius is large enough to lap over from one distinctive area to another. Bands of sky get processed in the same way as an adjoining mountain range. That leaves a light band (or halo) in the sky where it gets inappropriately lightened. One way to combat this problem is to modify the radius. In Dynamic Photo HDR (DPH) that's easy to do and users are sometimes encouraged to make the radius as large as possible so that its effects are spread out over a larger area and, thus, attenuated. The problem with that, in turn, is that you lose some flexibility in fine-tuning your results.
A little thought should convince us that the way to combat this haloing behavior is to precede the toning operation with one that breaks up the entire image into discrete regions. By using some edge-finder algorithm or the like the photo is decomposed into a number (probably a large number) of polygons. Now each polygonal region can be tone-mapped separately and stopped at the polygon edge so that no haloing results in other regions. So far so obvious. It seems clear that the Halo-Matix algorithm of DPH does not do that. The secret of using the Halo-Matix algorithm is to manage the halos; you're not going to get rid of them completely. But the Details Enhancer algorithm of PX does seem to do that. It seems as though that algorithm finds natural lines or edges in the photo and treats the contained areas separately from each other. Here it appears as though PX DE is using the dark and light lines of my test sheet to help break the sheet up into separate regions. Since each region is processed separately we would expect each region to have a different look. We actually see this in the results. I don't know that everything I've said is correct but that's the way it appears to be. The drawback with PX is that there are occasions when PX should ignore the lines because they simply cross over an underlying large feature that should be treated in a uniform fashion. This is probably related to the problem of distinguishing high-frequency features from low-frequency features. The low-frequency features in this sheet consist of the two or three 'towers' of circles (and the steps of the stepped gradient). The high-frequency features are the little compartments carved out by the underlying black and white straight lines (and sometimes the lines themselves as you can see in the thicker lines).
That's probably as far as we can go in interpreting this particular result.
I tried to improve this performance on the part of PX but I was only able to do this:
Again, I had to set the strength to the lowest setting and turn 'Luminosity' down in order to get an acceptable result. The compartmentalization is almost gone (naturally, 'Strength' is at 0) but no combination of settings would allow me to get rid of all the flaring or pinching. This is as good as it's going to get and, in fact, aside from the flaring the separation of dark and light disks is acceptable and a bit more visible than it is on the original (which is the whole point). We've lost the highest gradient brightness (is it my imagination but do the algorithms on both these products do better with darker values?). There's no point in pursuing this; we should turn to the 'Tone Compressor' algorithm. Here it is.
This particular result isn't too bad, really. There's adequate separation in the left and right disks but, on the settings I chose (I couldn't find better), we've lost separation in the highlights.
Now let's compare this performance with Dynamic Photo HDR. DPH has seven algorithms and I processed the test sheet for each of them and I include them along with my settings (I tried to optimize each). Here's the 'Ultra Contrast' algorithm:
Pretty good although it lacks something for the brightest values. Compare this to Photomatix's 'Tone Compressor' version above. Here's the 'Photographic' version from DPH:
This is the best we've seen so far. There's good separation in all three disks and every step of the gradient is at least indicated. What we've noticed so far is that the straight lines don't have any effect on the processing of the underlying figures. No 'compartmentalization'. DPH seems to do a good job of finding the high-frequency shapes and ignoring them. Now for DPH's 'Human Eye' algorithm.
Best performance so far on the gradient but the settings I chose reduced separation in the darkest disks (they're only separated by 3% as I said above).
Let's see if the 'Halo-Matix' algorithm lives up to our expectations:
Wow. It sure does. The gradient is rendered clearly (notice that the darkest zone has been lightened). All the disks are clearly distinguished. No flaring or compartmentalization (but there isn't anything like that in any DPH algorithm). Over all this might be the best. It sure looks like the place to start when you're doing HDR work.
DPH first became notorious because of its 'Eye-Catching' algorithm. It produced a lot of the hot and colorful results of the early phase of HDR. What's also clear is that it performs really well in black and white:
Not as good, on the whole, as 'Halo-Matix' but reasonable well. To wrap things up here is 'Auto-Adaptive'.
Not too good on the gradient but the clearest separation of the disks on the left that we've seen so far.
What do we learn from all this? So far I would recommend Dynamic Photo HDR for several reasons. It works better overall. It doesn't suffer from some of the obvious problems of Photomatix (which I and others have identified as 'pinching' or 'flaring' and also 'compartmentalization'). The product is more responsive to the controls. It has MANY more features than Photomatix both in terms of available algorithms and light controls. And the price is about two thirds of Photomatix' price.
I want to emphasize that we've only looked at black and white test sheets but, still, HDR is an idea about brightness and not about color. Also one legitimate use of HDR is just to take the brightness layer from Lab versions of photos and process that before returning it. So although these tests are limited they point the way to some important uses.
Till next time,
Bob
0 comments:
Post a Comment