Defining the ideal fisheye FOV for 2-shots

to fit with four current stitching S/W ability

Subject

2-shots panography (a.k.a. Double Shot) unique features amongst other stitching techniques (i.e. those that need more that two images to be stitched) are twofold:

Unfortunately, not all the fisheye lens images allow true 2-shots stitching, as:

Both of these points are related to Angle of View value of the complete circular image. While it can be fully covered on the digital camera sensor, the "FOV" might well be too small. It is easy to find that 180 degrees of FOV are definitly not enough and some have claimed peremptorily that 185 degrees is probably sufficient...

To fully benefit from the 2-shots positive features that is listed above, the idea lying behind this paper is then to determine which minimum FOV is actually required;

In a short statement, the driving specification of the study is:

Recall: Background for the study

No suitable DSLR fisheye lens is currently available:-(

We have successfully used a recent P&S high-end camera (Nikon CoolPix P5100) fitted with a FC-E8 fisheye converter to produce (thanks to the adapter "shaving") true 2-shots panography. Despite the 190 deg useful FOV, the image quality is unfortunately so-so in part because of the mediocre definition of the 12.1 MPix (!) images that are recorded on such a small sensor...

All trials that we have done with 2-shots from short focal fisheye lenses on DSLR cameras have yielded poor or inconsistent results.

By using a tricky empirical manipulation of images that were shot with 10 mm fisheyes, we have demonstrated that composite images fully simulating real fisheye images with about 190 degrees, seems to be perfectly suited for 2-shots. There also seems that only a small margin of the FOV value then is spared.

Sadly, there is presently not one short-focal fisheye lens on the market the market that meets the specification of 190 degrees for the usefull angle of view:

This was also the case for the older Sigma 8 mm f4.0 Canon 7.5 mm f5.6, Minolta Rokkor 7.5 mm f4.0 and Nikkor 8 mm f2.8 that are all no longer in production.

Please note that two Nikkor lenses had a FOV of 220 degrees: the monster 6 mm f2.8 and the 6 mm f 5.6. But both of them aren't suitable or affordable for digital panography even if you find one these days....

Description of the study

2-shots test sample images

It's possible to evade the problem caused by the lack of suitable fisheye lens: an ideal but artificial set of 2-shots images can be produced easily and these images shall have any angle of view value as you like!

The trick: From an existing equirectangular panoramic source image, PTGui can extract and convert these two circular images:

  1. By setting two yaw parameter values separated by 180 degrees (Images Parameters tab window) successively for each image of a pair.
  2. By setting the required angle of view (e.g. 185° x 180°) and selecting "Circular" (Panorama Settings tab windows).

This panography of the interior of a church was selected to perform the test:

Baziège (France).

The original is 7000 x 3500 pixels in its dimensions and was stitched from 4-shots. EOS 5D + Tokina 10-17 @ 10 mm (-2, 0, +2EV Enfused)

( 7+1) Sets of pairs of images of diferent Angle of View

Two images separated 180° in yaw (to make a pair of opposite views) were extracted from the equirectangular image hereabove and successively for the following FOV settings:

  1. 182.5°
  2. 185°
  3. 190°
  4. 195°
  5. 200°
  6. 210°
  7. 220°

Furthermore: two more opposites images with FOV of 190 degrees were inserted in a black rectangle simulating the sensor of a DSLR: the fisheye circle was willingly off-center both in x-axis and y-axis direction. Thus, it is the 8th pair of images that composed the whole set of 16 test images.

The circle is off-center by 200 and 30 pixels along x- and y-axis respectively.

The software

Four programs were used for assessement of the "optimum" Angle of View that should happen for the fiheye image. They are:

The first three used software are registered versions on the Mac. Panoweaver 5 is a demo version on Windows (v 4.0 used to be supported on MacOS but is now obsolete and awfully slow).

Sorry, being newbie, I could not manage to tame the Hugin 0.7.0 version on my Mac in time for this writing. Plan for inclusion of this Open Source program ASAP in a next to come revision of this paper.

The test

We have ran the four programs with each of the eight pairs of test images and a specific sequence was performed for each of them. The following describes approximately this sequence in main steps:
  1. Detection of the circle of image
  2. Settings the appropriate inputs and outputs data.
  3. Run full automatic detection of controlpoints.
  4. Previewing of the output equirectangular panorama (when available).
  5. Rendering of the panorama (when available).
  6. Saving of the output (draft) equirectangular file,
  7. Saving of the project file (at the outcome of the auto-detection step).
  8. Manual control to improve the draft version obtained at the end of the early sequence above:
    - Creation of additional CPs as needed to possibly better spread the clouds of CPs around the circle.
    - alternatively on APP only (as there is not yet manual creation/adjustement of CP implemented in v. 1.4.2): "Semi-automatic" detection/creation of additional CPs with the same intention.
    - Optimization of the panorama.
    - Rendering of the final output panorama.
  9. Saving of the output equirectangular panorama file.
  10. Saving of the project file of the final output panorama.

Results

1- Autodetection

Auto-detection of Control Points and observation of the panorama preview.

AutoPano Pro
PTGui Pro
Stitcher Unlimited-DS
Panoweaver
182.5°
Total detection failure
Total detection failure
Nice detection

Mediocre stitch

Fair detection

Mediocre stitch (PB zenith)

185°
Good detection Good stitch
Total detection failure

Good stitch

Nice detection
Good detection

Good stitch

190° off-center
Good detection

Bad stitch

Good detection

Good stitch

Good detection

Good stitch

Good detection

Good stitch

190°
Good detection

Good stitch

Good detection

Good stitch

Good detection

Good stitch

Good detection

Good stitch

195°
Good detection

Good stitch

Too high Pitch

Good detection
Much too slow detection
Good detection

Bad stitch (wrong FOV)

200°
Good detection

Good stitch

Good detection

Good stitch

Much too slow detection
Good detection

Very bad stitch (wrong FOV)

210°
Good detection

Good stitch

Good detection

Good stitch

Crash
Good detection

Very bad stitch (wrong FOV)

220°
Good detection

Good stitch

Good detection

Good stitch

Crash
Good detection

Worst stitch (wrong FOV)

2- Tweaking stage (manual)

Correction of Control Point misplacement, creation of new CPs to better spread the clouds of CP, optimization and final panorama rendering.

These operation aren't fully consistent with the 2-shots fundamental philosophy as full automatic operation should be the rule. We have added this step just to possibly improve the panorama quality by additionnal steps under manual control of the user.

AutoPano Pro
PTGui Pro
Stitcher Unlimited-DS
Panoweaver
182.5°
Total failure

Lack of true manual creation

Excellent stitch

(laborious process)

Good stitch

(laborious process)

Good stitch

(laborious process)

185°
Excellent stitch

(easy process)

Excellent stitch

(laborious process)

Good stitch
Excellent stitch
190° off-center Slighly deficient stitch
Excellent stitch
Good stitch
Excellent stitch
190°
Excellent stitch
Excellent stitch
Good stitch
Excellent stitch
195°
Excellent stitch
Excellent stitch

(easy process)

Not tried
Failure to enter the correct overlap setting

Very bad stitch

200°
Excellent stitch
Excellent stitch
Not tried
Failure to enter the correct overlap setting

Very bad stitch

210°
Excellent stitch
Excellent stitch
?
Not tried
220°
Excellent stitch
Excellent stitch
?
Not tried

Conclusion

The ideal minimum Angle of View appears to be somewhere between 185 and 190 degrees. Then all software have the capacity for detecting control points and to spread them sufficiently.

The sample images were here artificially evading all usual defects that affect the outer part of the cirular fisheye real images: no blurry edge, no bluish vignetting, no light fall off, ideal equisolid angle projection, etc. Therefore an ample margin should be added to the above figure to be more representative of the real world.

190 degrees would therefore be probably the best figure for specification of minimum Angle of View of a fisheye lens for 2-shots panography.

Michel Thoby

3 June 2008