Lightroom Pano Merge

[mcbroomf]

Sorry, I misunderstood your post the first time.  You're quite correct: the roof edge with the railing in the center foreground is not cleanly stitched.  I don't know whether that is the result of my sloppy handheld technique (I was maneuvering around a stream of other tourists as I made the individual shots) or some failing of Lightroom.

Yes, that's it.  I'm also suspicious of the railing at the end on the right.  It appears to end in thin air.

I wonder if you have too many overlaps as you mentioned.  It may be worth looking at the files and pulling one or more out in those spots if that's the case.  I don't know about PTGUI but Auto Pano is available as a 30 day trial and may be worth playing with if you have any plans to do something with the image.

[Chris Kern]

I wonder if you have too many overlaps as you mentioned.  It may be worth looking at the files and pulling one or more out in those spots if that's the case.  I don't know about PTGUI but Auto Pano is available as a 30 day trial and may be worth playing with if you have any plans to do something with the image.

I'm not sure how much time I want to invest in that pano—I actually like the skyline pano better as an image—but both of your suggested approaches are worth trying when I get some time.

By the way, I just checked, and PTGui also has a free trial.

[Paul2660]

I forgot to mention, that to my eye, I like the effect as it really creates a flow to the image.  The building are nicely rendered, straight lines to the curves of the tracks and railing.
Paul

[Chris Kern]

I forgot to mention, that to my eye, I like the effect as it really creates a flow to the image.  The building are nicely rendered, straight lines to the curves of the tracks and railing.

Interesting you should mention that.  I've encountered a number of instances—to my eye, at least—where some geometric distortion actually improves the appeal of an image.  For example, I've found that if I correct all the keystoning in a shot of a tall building, I lose some of the sense of its height.  I often like the photo better if I allow the top to converge just a little.

[Chris Kern]

For those of you following this thread who are interested, I've attached full-resolution crops of the two stitching errors Mike Broomfield pointed out.

[Paul2660]

The 2nd one is a bit more critical.  The first one if I am seeing the right spot is an easy fix. 

The 2nd one would mean grabbing the correct part from one of the images, copying and pasting into the image, then warping that selection.  Many times, you can fix such issues, can't tell on this one.  If not, it's rework from the start.

Paul

[Bart_van_der_Wolf]

The 2nd one is a bit more critical.  The first one if I am seeing the right spot is an easy fix.  

The 2nd one would mean grabbing the correct part from one of the images, copying and pasting into the image, then warping that selection.  Many times, you can fix such issues, can't tell on this one.  If not, it's rework from the start.

But there isn't much one can rework with, on an automatic stitch ... There are no controls in LR to do a better job, are there? Sure, postprocessing is always an option, but creating more work to correct such errors is not what stitching is supposed to do.

Cheers,
Bart

[Paul2660]

Personally my workflow is now LR first as I just prefer the dng output and ability to work the entire file as a raw. So if after a few attempts LR can't get it, depending on the error and scope I decide to either manually fix or move to ptgui and autopano.

Hopefully LR will improve over time but I doubt it will ever allow control point.

What a lot of folks don't remember is the old interactive CS3 option. I still fall back to that at times albeit only in 8 bit. That version allowed you to make changes to where rhe images were blended and many times a short tweak would fix a problem. And you could view the image at around 50 percent.

So farcical having very good success with LR.   LR is even able to work out a solution for my old 3 part stitches with the Zork adapter which in the past always vexed most software as the software did not like the lens camera combination. Canon and Mamiya 35mm.

Paul

[samueljohnchia]

Does anyone know what is the resampling algorithm used for panoramas in LR6?

[Paul2660]

Does anyone know what is the resampling algorithm used for panoramas in LR6?

That is a good question, which I have wondered also, both with LR and CC.  One aspect of Ptgui I have always liked is the ability to pick the algorithm for the warping. 

I have assumed that LR uses the same algorithm that is used with standard uprez during export. 

Paul

[dreed]

Fact is, that a properly executed stitch with e.g. PTGUI, will potentially deliver a higher image quality. That's a benefit in my book, but it's less convenient to get there from a workflow perspective. The hugely larger level of control of a dedicated stitching program, and the better resampling quality, can deliver much higher image quality and make some types of stitches possible that will fail in LR.

I've shot what I thought were easy panoramas that LR could not do by PTGUI did without a sweat. Go figure.

[samueljohnchia]

That is a good question, which I have wondered also, both with LR and CC.  One aspect of Ptgui I have always liked is the ability to pick the algorithm for the warping. 

I have assumed that LR uses the same algorithm that is used with standard uprez during export. 

Paul

It is getting harder to tell what resampling algorithm is being used.

LR6 uses a different one than Bicubic Sharper for downsizing images, as far as I can tell no one yet figured what it really is.

Photoshop has been using bilinear (!) for lens corrections for quite a while too. I don't know if LR uses bilinear for lens corrections. Changing the default interpolator in Photoshop's settings does nothing to force lens corrections to say use something like bicubic. That's why lens corrections always resulted in rather blurred detail in PS!

[Bart_van_der_Wolf]

It is getting harder to tell what resampling algorithm is being used.

Hi Samuel,

That's correct. Panorama stitching is often very much like distortion correction, some areas can get enlarged, others can get reduced in size, in the same image.

LR6 uses a different one than Bicubic Sharper for downsizing images, as far as I can tell no one yet figured what it really is.

Photoshop has been using bilinear (!) for lens corrections for quite a while too. I don't know if LR uses bilinear for lens corrections. Changing the default interpolator in Photoshop's settings does nothing to force lens corrections to say use something like bicubic. That's why lens corrections always resulted in rather blurred detail in PS!

LR uses a decent downsampling method (better than Bicubic and certainly better than Bicubic Sharper), although I do not know what it is exactly. The upsampling has some resemblance to bilinear. However, there may also be adaptive (to the image content) features at play. But the difficulty with stitching is the small rotations and enlargements/reductions. That will lose detail pretty quickly if the wrong method is used. Since LR's stitching is Raw input based, it's not easy to simulate with synthetic images which would show the trade-offs more quickly.

Dedicated stitchers like PTGUI could produce multiple conversions with the same pixel dimensions but interpolated/scaled with different filters, and the user can then blend the best parts manually, e.g. Lanczos2 for high contrast edges with sharp edge detail (yet with minimal halo), and a higher support version of Lanczos for structural detail, where ringing artifacts will be harder to see (if at all) due to the variable detail.

Cheers,
Bart

[samueljohnchia]

Hi Samuel,

That's correct. Panorama stitching is often very much like distortion correction, some areas can get enlarged, others can get reduced in size, in the same image.

LR uses a decent downsampling method (better than Bicubic and certainly better than Bicubic Sharper), although I do not know what it is exactly. The upsampling has some resemblance to bilinear. However, there may also be adaptive (to the image content) features at play. But the difficulty with stitching is the small rotations and enlargements/reductions. That will lose detail pretty quickly if the wrong method is used. Since LR's stitching is Raw input based, it's not easy to simulate with synthetic images which would show the trade-offs more quickly.

Dedicated stitchers like PTGUI could produce multiple conversions with the same pixel dimensions but interpolated/scaled with different filters, and the user can then blend the best parts manually, e.g. Lanczos2 for high contrast edges with sharp edge detail (yet with minimal halo), and a higher support version of Lanczos for structural detail, where ringing artifacts will be harder to see (if at all) due to the variable detail.

Cheers,
Bart

Hi Bart,

That's a nice way of looking at panorama stitching :-)

Yes, I recall reading your posts on downsampling interpolation in LR. How interesting that it is neither Bicubic nor Bicubic sharper. If I am not mistaken, in Photoshop, when Bicubic (automatic) is chosen, upsampling should use Bicubic Smoother. Strange that it resambles bilinear in LR!

I wonder if photographs of highly detailed targets could be captured, and then put through the paces of LR's stitching and PTGui, to reveal the trade-offs.

For my landscape pano stitches I am usually using Lanczos 8 in PTGui, with excellent results.

Interpolation to me is very important, I'm sad that it is not more vigorously discussed. I never saw it mentioned once in two recent lu-la articles on focus stacking. The whole point of that is to increase in-focus detail, so using an interpolator that makes everything softer is counter to the purpose. Unfortunately for the majority of the LR audience, a simple turn-key solution is preferable over fussing about such issues.

P.S. I had the opportunity to study at close range some billboard prints from Apple's 'Shot on iPhone 6' campaign, and they look surprisingly good. I wonder what upsampling software they used for those.

[Damon Lynch]

I'm intrigued to read so many have had good success with PTGui. Probably I need to put more work into learning it, because on all my recent files with respect to distortion the initial output from Autopano is so much more pleasing to me. The above image is from 37 files initially output from SNS-HDR, using a tripod. With this image LR HDR + pano combo produced notably worse output.

The below image is from more than 200 handheld source files. LR CC very quickly ran out of memory and failed to create a pano; Autopano chugged away for a couple of hours and did the job. As I recall in PTGui the initial rendering of the horizon was distorted (i.e. not straight). The uneven sky comes only from using a polarizing filter.



I had my tripod but I found rattling off the images handheld took far less time (helpful when you're on top of a hill with no toilet and a crook tummy ). Image quality subsequently does suffer in some places, but then again the air was so full of dust that it probably doesn't matter much in the overall scheme of things. Still some portions came out pretty well I think e.g.

[Bart_van_der_Wolf]

I'm intrigued to read so many have had good success with PTGui. Probably I need to put more work into learning it, because on all my recent files with respect to distortion the initial output from Autopano is so much more pleasing to me.

Hi Damon,

Autopano Giga is fine on large numbers of files if it actually works out automatically, however I find it cumbersome if manual adjustments need to be made. Also its colormanagement is a bit of a mystery to me. But it only uses a relatively limited number of projection methods, that are also available in PTGUI (and then some). So the perspective/projection distortion should look identical with the same projection selected in both stitchers.

The above image is from 37 files initially output from SNS-HDR, using a tripod.

Yes, that's also how I approach HDR scenery panos, first let SNS-HDR do its magic on the individual tiles, then stitch the tiles with e.g. PTGUI. Apparently you had too little space to back off a bit more, so you inevitably ended up with an extreme FOV which, in rectilinear projection, produces a somewhat stretched appearance.

The below image is from more than 200 handheld source files. LR CC very quickly ran out of memory and failed to create a pano; Autopano chugged away for a couple of hours and did the job. As I recall in PTGui the initial rendering of the horizon was distorted (i.e. not straight). The uneven sky comes only from using a polarizing filter.

In my experience, a distorted horizon can be simply solved in PTGUI by assigning a few horizontal control points to some of the images that are actually on the horizon. That would allow PTGUI to figure out the horizon's (Pitch) position in the overall scheme of things, and to level it at the same time. Polarization filters usually create problems is blue sky gradients on wide FOVs.

I had my tripod but I found rattling off the images handheld took far less time (helpful when you're on top of a hill with no toilet and a crook tummy ).

I can imagine. Distant scenes are relatively tolerant with stitching of handheld shots. Things get more critical with close up foreground detail or occlusions, and views through gates and doors/windows and such. Situations like you describe call for fast action, founded on experience. Practicing and experience does help.

Cheers,
Bart