Spherical panography

Panoweaver 4.0 Vs PTGui 6.0.3 Vs Stitcher 5.5.2-Unlimited/DS

Auto stitching of two hemispherical images:

Comparison test

17 December 2006

Introduction:

Early in November 2006, Realviz has introduced a new sub-variant of its Stitcher newer version software called -DS (for Double Shot) that may provide a 360 degree panorama from two opposite hemispherical fisheye images.

Since their inception, the Panorama Tools were able before to do so. Consequently the subsequent GUI (Hugin, PTAssembler, PTGui, PTMac to name a few of them...) that support PanoTools are also capable.

In the meantime, Panoweaver was introduced to specifically treat only Fisheye images, including of course two hemispherics.

This simplified panographical process is probably mostly used by the real estate -property- buy and sale trade. The customers for this specific segment of the panorama market are generally looking for simplicity and therefore the most automatic the production flow, the better it is for them.

Low resolution and low-to-medium quality of the panoramas have been keeping the more demanding panographers away from this technique. This drawback is now not fully valid anymore as some recent equipment that are very accessible and allow to give high quality 360 degree panorama output for the Web, from only two shots. I assume that more are to come on the market in the near future.

The three above mentioned commercial software have been lately updated with some new features and all three are now available both on Windows as well as on Mac OS X.

As long as the horizontal angle of view of the circle is more than 180 degrees, each is claimed to support stitching of two circular images and can render a spherical 360 degrees panorama. Of course one could possibly use up to three circular images (Panoweaver) or even more of these (PTGui, Stitcher 5.5.x-Universal). The following test refers only to auto-stitching two. IPIX was formerly the main provider in this field. Following bankruptcy the sale of its software has been recently discontinued.

PTGui and Stitcher (incl. DS) can explicitly correct geometric distortion by possibly using up two three separate correction parameters. Panoweaver may not do so and I suspect it uses only one correction parameter.

All the three stitching tools offer trial versions that can be downloaded from the respective websites.

The following is a report of a comparison test. That test was not exhaustive for all the features, but related to two important capabilities: sensitivity and distorsion correction.

Background: circular image and very wide angle of view

There are not to many lenses and cameras combos that can provide images having both full circular coverage and angle of view of much more than 180 degrees.

Some Nikon CoolPixes compact digital camera fitted with a fisheye converter (FC-E8 and FC-E9) are still available (mainly on the second hand market for the former). While these combinations have been the workhorse for many Real Estate panographers for a long while, I shall not consider these cases as the low resolution that is obtained make those tools rather outdated.

An outstanding lens (Nikon 6 mm f/2.8, 220 deg.) had been manufactured by Nikon but its production was discontinued. It was very rare on the market, very bulky, very heavy and very expensive. Coastal Optical Systems may still manufacture a similar monster (7.45 mm f/2.8, 185 deg.) and a more reasonable one (4.88 mm f/5.2, 185 deg.).

I don't know of another widely commercially available lens that has a specified angle of view well above 180 deg.

There has been an interesting discussion on this topic on the Stitcher forum recently.

While this is totally hidden by its designer, a recent fisheye lens presents such outstanding optical properties: the Nikkor 10.5 mm f/2.8.

Its actual circular image useful coverage is more than 190 degrees when the sun shield is "sawed off". When it is adapted on a fulframe Canon EOS 24 x 36 DSLR camera, it produced a "barrel type" image because the sensor width crops off some of the circle.

Nevertheless, this partially circular images may produce a complete 360 degree panorama horizontally. The vertical FOV is about 140 degree and thus outputing a cylindrical projection of unusually high VFOV or alternively a slightly incomplete spherical 360 x 180 panorama, is feasible.

Test objectives:

I have elected to compare only to characteristics from the stching software:

As it is rather straightforward to shoot such images with the EOS 5D + 10.5 mm combo, it is interesting to know if the autostich does work then with image having normal features for the software to auto-align.

Test #1: Sensitivity

I had tried a while ago to make a panorama from two images to wrap up another comparative test reporting (see § "Test number four" near the end of the paper). I had no success at that time.

Things have changed and Realviz has now introduced a version specifically aimed at this purpose, making two with EasyPano's Panoweaver.

I have therefore decided to revisit the whole case including PTGui 6.0.3.

Source images:

First part -A (3 pairs of images)

Taking into account the S/W version evolution, I have first made a repeat of the test that I have mentionned above and have add Panoweaver to the duo.

I then ran three successive tries with three pairs of images.

Six images had been shot at 60 degrees interval in Landscape mode. From these 6 images (# 1, 2, 3, 4, 4, 5 and 6) one set of 2 images (# 1, 4) has been selected and subjected to automatic stitching test:

Second part -B (1 pair of other images)

These last images duo were shot with intention to run the test next to come (w/ full circular images)

Results:

Sensitivity test
Test1 part #-A
Test1 part #-B
Panoweaver 4.0
OK
OK
PTGui 6.0.3

(Max 30 pairs of CP allowed by setting)

PTGui has found no Control Point at all.
PTGui has found only 8 pairs of control points among 30 possible (all 8 are on the same end of the images)

REM: more clicks on "Align panorama" give some additional CP's

Stitcher 5.5.2-U/DS
OK
OK

Conclusion (test #1 = sensitivity):

Test #2: Peripheral annular distortion correction (10,5 mm Nikkor adapted on 24 x36 DSLR Canon)

The Nikkor 10.5 mm lens is a real optical wonder. But its nice qualities have been obtained at the cost of quite unusual high geometrical distortions: the image mapping compression near the edges is much stronger than for other fisheye lens. You may read this article (see §2 "Image Mapping") to know more about this. The "defishing" part of the software workflow is therefore challenged for outputting regular and correct images in the course of the stitching process.

There is only an annular part of the first image that overlap with the similar area of the second source image. With only this information, the software should guess how to transform the images to put them in the equirectangular projection. The overlaping area is also the part of extreme compression. The rest of the image that is more central to the circle has no corresponding overlapping pixels and cannot provide useful information for the mapping and transformation. Extrapolation of the observed radial trend of compression has to be done and this is not trivial at all without prior calibration of the fisheye lens image mapping geometry.

Finally, in case of failure in filling any of the two requirements, the ability provided by the software to manually correct the possible flaw shall be also finally studied and evaluated.

Background: to complete a circular barrel image

As the image that is got from the combo that I presently own (10,5 mm Nikkor + Canon EOS 5D) is not totally circular, this hampers the sensitivity of PTGui to reliably detect matching Control Points. This was described in a precedent paragraph.

To be able to make a fair comparaison of the respective ability of the three contenders in correctly correcting the distortion, I had to find a workaround and I have thus designed and made a set-up that allows to bypass this shortcoming:

I have "simply" mixed portrait and landscape images (PS CS 2) of the same scene all accurately shot from the same No Parallax Point (with a common optical axis). The set-up principle is shown here:

These 2 circular composite images are the specimen that are to be stitched in the test:

Important: A special feature has purposedely been included on the Nadir of the scene. A dart circular printed target board was centered exactly on the Nadir point.

Any residual radial distortion could then be spotted at once:

Results:

Distortion correction test
100% Auto-stitch
Manual adjustment of distortion correction parameters
Panoweaver 4.0
Obvious residual annular compression in the overlap area besides important stitching errors
No manual correction means seems to be available to the user.
PTGui 6.0.3

(Max 30 pairs of CP allowed by setting)

1) Good result with "Medium" lens distortion minimization setting. "b" is being optimized alone. Some minor stitch errors.

2) Obvious residual annular compression in the overlap area when setting to "Heavy" or "Heavy + Lens shift"

Apparent "perfect" result when applying a, b c, d & e correction coefficients values gotten from a separate previous stitch calibration with 4 Portrait images.

Can make a good dedicated template file.

See the QTVR

Stitcher 5.5.2-U/DS
Obvious residual annular compression in the overlap area
I could not find a correct set of "parameters" (incl. circle radius) to perfectly correct the annular compression. (Chaotic behavior in response to small changes in ths settings)

Conclusion (test #2 = radial annular & peripheral distortion correction):

Stitcher 5.5.2-U/DS and Panoweaver do not automatically correct the unusual amount of radial compression that the Nikkor 10.5 mm fisheye imposes on the peripherical ring of the circular image.

PTGui may provide either good result or as bad result as the other contenders, depending on the source images and the initial sttings (Wild guess). I have also noticed that the output depends on possible repeat of the click on "Align images" button.

Panoweaver seems to give the worst result from the three when it is supposed to be a fisheye image dedicated performer, if I have well understood. But I must confess not being an expert user of this dedicated software.

Possibility of parametric adjustment:

PTGui gives excellent outputs when the settings have been adjusted with carefully calibrated parameters and correction coefficients.

I presume Stitcher could probably also be adjusted to do the same. Unfortunatelly, I was personnally not able to get good and fixed calibrated parameters (from 4 images stitching for instance) that would give correct stitching from two images. I have discovered a chaotic behavior from this S/W as successive tries with the same set of four images yield wildly different values for the parameters -yet the resulting panorama is generally acceptable! Complete disorder happens when any one this parameter set of values is put into the "Properties" windows dialog. Obviously the two Stitcher modes (ST-U and ST-U/DS) do not share the same algorithm.

BTW, I still don't know for sure what are the real significance of some STitcher 5.5.2-U essential figures (such as "correction parameters", "angle"..) that are shown in the "Properties" window and I could not tame their blunt and weird mutual interaction either.

I did not find a way to feed calibrated parameters values in the DS type mode of Panoweaver.

Overall conclusion:

The Nikkor 10.5 mm lens is one of the few lenses that can provide high quality / high resolution images that could be stitched in the "Double Shot" style (i.e. from two opposite quasi-hemispheric images with more than 180 deg. HFOV). It has then presently to be "adapted" on a Canon FF DLSR.

Unfortunatly, I had difficulties during this test to get consistant good results from any of the three software that support this process both under Windows or Mac OS X.

Panoweaver is very finely tuned to be very sensitive but it has been a poor performer in my hands: bad stitch and poor distortion correction was constantly output along my tests.

Stitcher-U/DS is "potentially potent" and is very sensitive. But I could not find the right calibration settings that could improve the bad correction of residual radial distortion from the Nikkor lens. I could not find the information or tutorial for that.

Ironically, only PTGui may be able to auto-stitch correctly these challenging images and especially if calibrated parameter values are set in a template. But it is affected by too low a sensitivity when it is given with some images and no Control Points are found... Manual processing is a good workaround but it nullifies the purpose of the DS way.

I then choose personnally to stay with other more classical ways to make my spherical panographies: Three images shall be the minimum, unless I find a very compelling reasons to do otherwise. This may then be another future story;-)

Michel Thoby

Updated 24 December 2006