Very short focal fisheye lenses
Stitching/Blending Double-Shot (DS) images

Part (3)

Summary of the study

Sigma and Sunex have recently and almost simultaneously put on the market a new very short focal lens. Both the Sigma 8-mm f2.8 and the Sunex Superfisheye 5.6-mm f5.6 may allow to do spherical panographies by assembling only two hemispherical opposite images (aka Double Shot or DS) thanks to their Angle of View >>180 degrees.
The noteworthy point is that this can be done now on an affordable basis unlike with the previously available but expensive alternatives (4.88-mm f5.2 and 7.45-mm f2.8) sold by Coastal Optical (also known as IPIX lenses) and not quoting the venerable (220 deg FOV) Nikon 6-mm f2.8 monster.

Part 1
Appraisal of the lenses performance shall be done in a pure practical approach without Chart measurement: I did not have the chance to put my hands on the lenses and had only some kindly provided sample images for experimentation.
While the Sunex is a fixed aperture model, the Sigma can be set from f/2.8 through f/22 and the capability of this lens shall be evaluated by taking this feature into account. The illustrated results are presented in Part 1.

Part 2
Evaluation of panorama stitching programs likely to allow DS stitching is presented in Part 2.

Part 3 (This page)
In the course of evaluation of possible ways to produce panorama with DS, I have discovered an anomaly
I have thus elected to describe the conditions where a strange dysfunction of some blending engines happens seemingly on a random way, when they are faced to blending DS images.
After many hour of experimentation, I have yet not found a valid explanation for this instability and I have resigned myself to write Part 3 (This page) in order to give all means to replicate for further investigation by others. I hope to be able to remove this part as soon as the problem is solved.

Note: While the source images from the Superfisheye where provided by Sunex, inc. differnt sources including Sunex he source images shot with the Sigma 4.5-mm were all kindly provided by Roberto Gomez Torres.

Part three

What went wrong -sometimes- when blending DS?

The problem:

While I was doing extensive experimentation of stitching program with Double Shot images, I have repetitively yet randomly stumbled against a strange phenomenon. I did not pay too much attention at first as I then believed it was resulting from an operating error from my part: the blender engine seemed to behave wrongly under mysterious circumstances.

At first, because of the randomness of the problem occurrence I hadn't realized that Enblend and PTGui internal blender engines only were concerned by the anomaly. I had first wrongly suspected the software selected for warping then I have also wrongly suspected a problem in the blending of +/- 180 degrees seams (not totally discarded though) and I had to draw the table here above in Part (1) to arrive finally at the conclusion presented hereafter.

In short: The blender seems to have been totally inactivated (or sometimes inefficient) while blending the output (DS only) panorama.
Just changing very slightly a parameter value setting (e.g. circular cropping size or location, by one pixel) may make the phenomenon to appear or disappear.

This instability lead me to make uncorrected conclusion until I finally could pin down the factor that symptomizes the problem (or that I think it does).

I guess the best way to describe it is by images and VR...


You may right-click (ctrl-click on Mac) on these thumbnail images to download the two full scale JPEG images for test replication.

Note: the original images have been cropped to a square (well, almost: 2293 x 2294 pixels) visually circumscribing the image circle. I had noticed that it was very difficult or impossible to correct the "vignetting" with ACR 4 on the rectangular uncropped images: the correction tool works by assuming the worst black vignetted part of the "usefull" image to be in the corner of the rectangle. This assumption is only valid for standard rectilinear image or full frame fisheye, not for circular fisheye images, of course.

Some of the outputs and corresponding project script files

A- Blending without the problem on the output panorama

PTGui script file

Panorama output with Enblend from a "good" script file: CubicQTVR 900kB

B- Blending with the problem on the output panorama

PTGui script file

Panorama output with Enblend from a "bad" script file: CubicQTVR 900kB

Panorama output with Smartblend from the same "bad" script file: CubicQTVR 900kB

This zip file contains the source images files and the project script files quoted in this paragraph.

You may replicate the blending with these project but beware to NOT optimize before Creating the Panorama even if PTGui ask you to do this. If you do the output might not show the intended information about the anomaly.

C- What are the differences between "good" and "bad" project script files?

There are not too much differences: two pixels in the dimensions of the cropped circle in accordance with 0.07 degree of FOV. Even so the output panoramas don't look the same!

D- A similar anomaly with the PTGui blending engine?

With the same source images as above and with this project script, one can get this output panorama (cubicQTVR 1MB).

A thin black line on the seam indicate probably a poor blend of the same problematic seam that plagued the previous case with Enblend.

Some more investigation might be done on this.

Michel Thoby

20 February 2008