Thursday, December 25, 2014

Hektisch, a panoramic story…

One word: Hectic. The week was so in fact, but this word describes more precisely the panorama that could not just give some easy time to be done. Been three weeks trying to fix this issue, and I won't be mistaken if I simply called a "panoramic adventure!" However, it was a case worth of a study, thus I'm hoping to dedicate this post for this panorama alone. I've made 3 trials already with this panorama and despite the hardships I'm having about it, I don't think I will be going there again to do it a 4th time!

The Beginning

The story begins not the in the current time. The idea of doing this panorama was actually planted in my head long ago before I even travel to Ireland at the end of September. Many factors were holding me down, but the number one factor was: Summer. The second factor was the security guys around the place. This place, Soug Sharg (Sharg Market) was specifically a scene of some confrontations in the past with security guys; day and night. Doing a panorama close to that place (and I guess belonging to it) might get the beehive disturbed. These two main factors held me back, and the rest was left for the proper timing and the capabilities of the body.
After coming back from Ireland in mid October I got busy with many things, and I need some time to settle my mind on how to do this panorama. I paid first a visit; just a plain visit without anything in hand, but simply walking around the place. Of course, the visit was nocturnal by nature, as this place actually shows more beauty at night with the "lantern" of lights glowing then. I did simple shots of this location long time when I was a beginner in photography and I did indeed, back then, made a simple vertical panorama for the face of this corridor in the outdoors. This time, however, the story was different.

The location under inspection,
right behind Soug Sharg's Marina.
The difference comes in many aspects: the time, the size, the intent, the procedure (the erratic part of the whole story). In this plain visit I had to check if there is a chance to do a centralized panorama in this place. The answer was positive; there is a distinctive center for the place. The lights in the place would provide sufficient light for simple exposure timing in case of catching up with the architecture of the place. However, long exposure is needed if the sky is to be given some color and not keeping it simply dark. Hence, HDR comes to the mind naturally. However, for the factors I've mentioned before (i.e. security people) I had to think thoroughly to work as fast as possible and hence "real" long exposures were not an option naturally, neither metering the light on location to find out a proper intermediate value (and it won't work because the difference in luminance is far greater than 2 stops).

Versuch Eins

The first trial was a failure by nature. I went out too late to the location thinking that too late is better to ensure not much activity around (people or security). One important point was forgotten here: lights are automated, and they switch off around 4 a.m.; Ironically, this is something I used to know from last winter's experiments with long exposures on the beach but my mind was completely absent.
The result was, a short trip, to say the best. In fact, I spent a significant time trying to centralize the set to make it in the center as much as possible. With places like this, symmetry is of the optimum importance and I had enough already with my left-handedness temptations, where everything looking normal to me with the naked eye is, in fact, tilted or skewed to some direction and not perfectly central. For this, my work was greatly done with the LCD of the camera. Shifting the tripod around was not an easy job too - moving a heavy set just for few centimeters with some delicacy is never easy, specially with the tripod been extended to its full length, as I was trying to have a slightly upper view than the normal level of the eyes.
After doing all of that and all this back-tiring work, I started to shoot. But first, I had to make one shot using the white balance disk to set it as a reference point for the camera's white balance point. The work then started and as soon as I reached the 4th angle in the process (and that would be at 120o), the lights simply turned off, leaving me in the dark. It was a mixture of anger and disappointment, naturally. But I guess I can blame myself because I should have been more attentive to such matters from previous experience.
Settings used here were f/9 (while using the hyperfocus principle), and bracketing (of exposures) to merge as HDR slides later.Metering here was set to Evaluative, which is meant to measure the light from different areas in the scene and then evaluate an average for the different zones of the scene, and naturally some exposures took 30 seconds already (because of the camera limit to this time in Av mode, otherwise it should have been longer).

Versuch Zwei

The second trial was set the week after the first. This time I've decided to head earlier than the previous time, and that was around 1 a.m.. Again, centralizing was essential and did take some time to do. The white balance was still saved from the last trial and the aperture too; all I had to do is shoot. This time, however, the tripod was lowered to around the chest level for easy control and viewing. No nadir (bottom) shot was taken as I would need extra work and careful planning for such a shot so I had to neglect it despite the fact that it bears a significant detail in the tiles underneath the tripod. I thought that I will manage to get over such a problem later. Almost same settings were used, except that metering was set to Center-Weight (thinking that the light areas are greater anyway from those dark areas in between).
Back home the surprise was on its way to shock me. It is regular practice for me now to work with modeling; i.e. using simple JPEG images to stitch a correct model of the panorama and apply the mold of this model on the HDR panorama, since the stitching program (as of lately) doesn't recognize much details from HDR images. To my surprise, the computer made a pizza out of my panorama!

The pizza made by the computer trying to stitch my panorama

I did not really understand what was going on here. Why did that happen? Did I miss some angles while shooting? Even optimization (i.e. correcting control points) did not work on fixing this "Pizza". I even tried to use Black and White images to make the model but the same issue occurred - a pizza! The thing is there was no problem generating control points at all (except for one or two slides beside the zenith slide and this is normal).
Investigating further down the road of making a panorama, the natural culprit to be suspected first here is the rotation axis, which if not picked out right, can cause the parallax error - an error that is a nightmare for panorama makers. Checking on some numerical values with a little research about the matter, I've found out that my camera and lens combination was indeed under a shift - This is my first full-panorama to be done after coming back from Ireland and apparently the VR-Head had some screws loose and there was a shift in the arms which I didn't check thoroughly.
Checking the images on computer to see if there is really any significant shift in virtual spaces between objects in the scene (when camera is rotated) I did indeed notice a really tiny amount of shift and change (e.g. the small green barricades and their background had some relative motion with respect to each other), but I was skeptical about the seriousness of this shift - Was really enough to do such a disgusting shock to my eyes, and even squander the computer's mind to the limit of losing control in that way?!!! Anyway, the shift of the rotation axis was fixed according to the numerical values I've obtained online and I was ready for a third trial.

Versuch Drei

For this trial, I was already tired of going to that place (and it had a non-pleasant stench to it), thus all what I picked was my tripod with the VR-Head, and my camera. This time I didn't wait for the weekend to come, but I did it in the middle of the week as my sleeping pattern was disturbed (after some headache issues which forced me to sleep afternoons!). All settings were already set (including the white balance from the first trial 3 weeks ago before this third trial). This time though I've returned to the Evaluative metering.
Back home I've converted the images directly to Black and White slides (thinking that colors can add to the confusion in connecting and stitching the images of the panorama) but again, to my surprise, the modeling process didn't work and just like in the second trial, the computer made a pizza out of my panorama! And again, the optimization process did not yield any useful result.

My third trial's pizza!

At this point I've realized that the shift in the rotation axis was not the only problem in paly here. There was something mysterious going on!
In a desperate move here, I've tried to apply templates (i.e. models) from other panoramas not related to this one at all. Specifically, I've tried to apply a model file for a panorama that I've somewhat faced some problems like this with: an taobh istigh an Waterfront (inside the Waterfront).

An taobh istigh an Waterfront

I have to say it was the first time I try such an approach; applying a model of completely unrelated panorama to another. Yet, the results were significant! However, the stitching errors did prevail and were hard to fix too.

Some of obvious stitching errors (and some were cropped from below) after trying to apply models not related to the same panorama itself. Just to note, this approach was done with the panorama done in the second try.

At least for this point here, I was able to discover the other projections and how well would they do and the results were astonishing (but not to be published now). This made me more eager to solve the matter of this mess. Despite the promising look of this approach but the main problem is that optimization cannot be done to correct for such stitching errors, and those stitching errors would need significant time to be fixed, if they were flexible enough to be fixed anyway!

Versuch Vier

At this point I was tired of the situation and really, really did not think of going further. There must be something related to this issue other than a simple physical attribute such as a shift in the no-parallax point and axis of rotation. Thus, my 4th trial now is to be done completely digital and on computer.
Despite the stress (yes, it was a real stress) the accompanies such unexplainable results when a panorama-making process goes erratic for no obvious reasons, I tried my best to organize my thought and work.

In the beginning, a general check up for the whole set of images yielded a fact that there is no problem in generating control points, and yet, optimization gives the "very bad" indicator. This usually means really bad control points (points connecting between two images) - that is, some points do not match or simply very distantly related. Anyway, after deleting these bad points, some images were left with no control points at all. In other words, complete images here were a major reason for the whole upset in the "pizza"! Now what to do?

First approach was to divide the work: dividing the panorama to 3 zones and work on the control points for each. I've reloaded the images anew, and canceled the repeated shots (those at 0o and 360o, keeping just one of them instead of the two). Then, I've generated (manually) control points between each 2 images in the first set of images which compose the middle zone of the panorama. At this point, optimization for this set of images alone was "good" or "very good" and them middle zone virtually out of stitching errors so far!

The same approach was done for the lower and the upper zone of the panorama (excluding the zenith point) with varied degrees of optimizations, but all were good so far! The tricky part now was in connecting these 3 zones together, as the panorama now was divided to 3 perfectly-matched zones! Thus, I've went on generating control points (again, manually) between vertically overlapping images in between the middle and the lower zones, and then again from the middle and the upper zones. It was tricky here as when optimizing any two sets of images after adding the control points, the third set gets corrupted and out of sync. Needless to say here there was a great deal of work regarding other options and properties needed to be controlled, mainly the Blending ratio for some slides in the lower zone. The situation was settled when I optimized the whole set of images from all zones; optimization then yielded "not so bad" - which was good for my purpose here at least and the three zones were matched with minimum of stitching errors. The zenith slide (topmost) was then matched with some images from the upper zone and optimized alone. Things look relatively perfect now except for few minor stitching errors AND some smudged zones resulting from luminance problems when merging into HDR (which is a resultant related to the exposures making up the HDR itself). Those are not a big nuisance for the time being but I guess they can be dealt with later.
Anyway, there was a minor problem of keeping the pillars at vertical lines which proved to be problematic. The beauty of the place is in fact in the vertical straight lines and such twist and tilt in those pillars is not a good compromise, so to say.

A model stitch. Notice the tilt of the pillars; some are slight and some are strong. Of course beside the other stitching errors.

This convergence of vertical line does effect the visual strength of the panorama as well as the ability to centralize the panorama and achieve a good symmetry (remember, the main reason for picking this location is the symmetry!). Trying to fix this problem with the usual approach (changing the roll and pitch angles) did not yield a proper solution to this problem. Thus, I've decided to go through the middle zone and assign vertical control points and some horizontal control points. These are special type of control points that tell the computer that such lines are (and should) be perfectly vertical or horizontal, of course depending on specific features in the scene. Thus, most of the pillars in the middle set of images was a target to this addition of control points, and optimization was done once again to apply these changes. The results were a joy to my eyes!

A final test stitch. The arrows point to the stitching errors and smudges caused by HDR problems. Compare the tilt of the pillars here with the black and white model above which was made before adding vertical and horizontal control points.

All what is left now is to apply this model to the REAL thing: the HDR set of images itself. After that, there will be a long journey through the different projection styles and the different tone-mapping (and each would need some degree of fixes of course).
I can feel the steam venting out of my brain. I've finally settled down with 3 weeks of work-failure-work relationship here! I have to say, I do consider it an adventure nevertheless.

Thoughts and Conclusion

The question now remains: Why did that happen? What went wrong here? Despite the simplicity of the situation, yet this panorama persisted in its stitching errors (even with the final results), and despite the reality that there was indeed a shift in the rotation axis away from the no-parallax point of the lens (Canon EF 15mm f/2.8 fisheye), yet fixing such problem did not bring any practical enhancement to the situation. This made me think of some points here that might have caused this problem and some of them if not all are kind of uncontrollable (to me at least):
  • Moving shadows. Because of the abundance of light sources on location there were multiple shadows in the place beside my shadow of course, which was moving and probably did appear in more than one slide. 
  • The symmetry of the place could be a culprit despite being the reason sought for creating this panorama. The resemblance between two images which are NOT related to each other could have possibly (just possibly) trick the computer on thinking that these two images ARE related. Hence, control points were generated to two unrelated images making a pizza out of the situation!
  • Just like the moving shadows, reflections on the ground might have a minor role here in "baffling" the computer and tricking it into generating what I might call False Control Points
  • By now, it is very common that repetitive patterns in the space do undoubtedly produce stitching errors; Galore! Thus, I was not really surprised to see stitching errors in the lower zone of the panorama where the tiles make a lot of lines crossing each other.


Well, after this long story (and hopefully not so boring one), I'm just glad and happy to see such an achievement. I didn't get such a feeling of accomplishment for a while (even though it wasn't far back behind when one of my panoramas was displayed in a luxurious coffee table book!).
I dedicate this post to "some" people who thought the work of photography is simple a deed of Photoshop and playing with the camera, specially when it comes to panoramas. I just hope this post would open the eyes to the hassles that we, photographers, must put up with just to come out with something - and most of the time, it is for our own satisfaction and no one and nothing else. For those work already with this venture of photography and reading this, I hope it is of some benefits to you all out there. There are some details that I couldn't mention here of course, but I'm sure you can figure them out on your own!

Merry Christmas to everyone celebrating it, 
or as the Irish say: Nollaig Shona Dhaoibh.

No comments:

Post a Comment