What is the ultimate quality with 2-shots panorama?

Let us compare with the standard 4-shots...

Version fran├žaise

Subject

1990's

2-shots was a common way to make quick and dirty panoramas in the 1990's and early 2000's. To make such a panorama only two shots were needed to produce the single pair of two hemispherical circular source images.

In that era, most of these panoramas were the production of photographers working for Real Estate businesses but not only. Most if not all used a P&S camera fitted with a fisheye converter. The resolution of the camera was then quite modest (at about 1 to 3-5 Mpix) and the quality of the output was seriously hampered by the poor appearance of the seam between the two images:

Most of the professionals and passionate amateur panographers aiming at higher quality used more than 2 sole images to stitch They were using standard lenses or fisheye lenses. They produced cylindrical QTVR panorama then some time later full spherical panorama.

Today

This separation is still valid: 2-shots are mostly done by Real Estate panographers and the quality of the panoramas is still generally considered mediocre even if the resolution has slightly increased by means of more pixel density on the modern P&S camera sensors.

The serious panographers still disdain to make 2-shots. Their DSLR camera can be fitted with a fisheye, but they shoot more than two sources images per panorama. They may often use a panohead to make multi-rows panorama and so get higher resolution in the output image.

Additionally, only few affordable fisheye lenses that could possibly help producing 2-shots panorama with DSLR are available. Unfortunately, I have found that their angle of view is much too small to overcome the problems listed above.

All these facts have contributed to continue the bad reputation of 2-shots while the quality of the camera has attained remarkable quality.

Question:

What is the quality that could be obtained with two-shots and today's technologies? This is what I shall try to demonstrate in this paper.

But let's us first recall why 2-shots is unique:

Reminder: 2-shots is the only way to avoid parallax errors completely when using fisheye lenses.

It is now well known that the NPP (No-parallax point) of a fisheye lens is not fixed as it depends on the aspect angle between the line linking the object of interest to the NPP and the longitudinal axis of the lens. The NPP may be located for example on a region of about 18 mm on this main axis. Situated near the front apex of the lens (for an angle of 90 degrees), it shall be at the level of the golden ring 15 mm behind for 45 degrees. The direct consequence of this optical intrinsic property of the fisheye is that to be totally free of parallax error, the seam that joins two consecutive images in the output panorama has therefore to come from a constant aspect angle all along this seam.

There is only one way to get this: stitch two opposite full hemispherical images.

Fortunately, most of the time the subjects are not very close to the camera. Then the stitch errors due to parallax when assembling more than 2-shots is very small and become invisible if the corresponding NPP is set correctly. But if an object is very close (from some centimeters to some decimeters) and on the overlap between consecutive images, there shall be parallax and subsequent visible stitch errors.


The reachable level of quality from 2-shots

The theory:

The NPP for 2-shots has to be very precisely located to fully benefit from the intrinsic property. But when this carefully respected, even if close-up is done during shooting, no worry about parallax errors are to be feared from.

When this prerequisite is met, two main factors still impacts the quality:

The transverse chromatic aberration (TCA a.k.a. lateral color) and the distortion can generally be well corrected by most dedicated software as long as the radial mapping of the fisheye is not too much different from the optical canon: the equi-solid angle projection is generally the standard for the software.

I have performed exhaustive and numerous tests to assess the power of the 2-shots technique. I have then found that the minimum Angle of view (FOV) that must be covered by the camera sensor is about 188 degrees excluding the bluish and blurry ring that surround the image circle. Furthermore, the greater the diameter of the image circle, the better can be the quality (of course).

The ideal fisheye lens is yet to be produced:-(

No question, due to sensor size, the best affordable camera for making excellent quality panoramas are IMO of the DSLR type. But there are no affordable fisheye lens presently available on the market that would meet the FOV requirement and this is a pity: There is no ~8 mm fisheye with FOV at least 188 degrees (for FF DSLR) and no ~6 mm also with FOV at least 188 degrees (for APS-C or 4/3 DSLR) on the market yet.

The best suited camera:

As the fulframe 24 x 36 sensors are obviously the best probable candidate to produce the ultimate quality, I shall then consider this type of sensor for the demonstration. The APS-C sensor may produce good images but they cannot rival the FF when all other features are kept equal. The surface of the full frame sensor photo-sites is double the area of the APS-C sensor and this make the difference.

Today (June 2008) the ultimate resolution can be obtained from the Canon EOS 1Ds Mark III with 21.1 MPixels. The diameter of the possible image circle (FOV 190 deg) shall therefore be about 3740 pixels. Sony has announced a 24.6 MPixels CMOS to appear in late 2008: then the circle diameter could reach 4030 pixels before cropping by the sensor width.

I personally use a Canon EOS 5D and a Tokina 10-17 mm zoom fisheye lens that was shaved from its sun hood to fit nicely with the camera. If the focal is set at 10 mm, the image circle is 3700 pixels (It could be 4200 pixels if the focal of the zoom is set at about 12 mm). But this circle is cropped along the two long side of the sensor while the FOV of the circle is >190 degrees.


The empirical demonstration:

2-shots circular images

We can therefore simulate exactly the uncropped circular image by overlapping two orthogonal images. These two images are shots respectively in portrait and landscape mode while the optical axis is kept at the very same position and the NPP is also kept still.

Landscape mode
Portrait mode
Composite full circular image
View looking toward South

Another pair of images is shot in the opposite direction to complete the 4*pi steradians full sphere with the previous pair.

Landscape mode
Portrait mode
Composite full circular image
View looking toward North

There are several way to assemble and stitch those images. But in any case the best output that we can get from the two pairs of orthogonal images from the Tokina is the exact simulation of the panorama that could be got from a perfect hypothetical 7.5 mm fisheye (190 deg FOV) on a 1Ds III camera....

The four source images produce no less but also not more useful pixels (when the redundant overlapping pixels are discarded) that what would be used for stitching of two true hemispherical 2-shots image!

I have used PTGui Pro to manually optimize and stitch these images. The output panorama is reduced from 7000 x 3500 to 5000 x 2500 before QTVR conversion.


The 4-shots reference case

The same panorama as before but using only four images in portrait mode.
Looking West
Looking North
Looking East
Looking South
Portrait mode

I have used PTGui Pro to manually optimize and stitch these 4 images. The output panorama is reduced from 7000 x 3500 to 5000 x 2500 before QTVR conversion.


Conclusion:

A 2-shots (a.k.a. Double Shots) panorama doesn't fully equal the quality of the reference 4 (or +) shots panorama.
When it's scrutinised at high zoom-in level, the join at the seam between the two image is slightly less sharp than in the reference 4-shots case and some color artifacts can sometimes be seen. On the Zenith and Nadir regions both methods give the same quality.
But even for Full Screen viewing on the web, the difference can't hardly be seen at normal zoom level.

The bad reputation of blurriness and color fringing that plagues 2-shots is IMO largely over emphasised today. It's due to the memory of the early years of Panography when the technology was still in infancy.

This reputation of mediocrity is unduly enduring simply because of the absence of adequate fisheye lens to fit with modern DSLR cameras. No one have really viewed a "modern" 2-shots panorama, because there aren't: only P&S cameras adapted with fisheye converters are available for doing 2-shots.

It 's a pity that none of the affordable professional optical products available on the present market features a large enough Angle of View (i.e. 190┬░) and subsequent image quality on the edges to fully benefit from the unique property of 2-shots.

As of mid 2008, the assurance of total absence of parallax error and the subsequent ability to mass produce high quality panoramas (notably but not exclusively for the real estate business) is sadly a fallow land:-(


Michel Thoby

26 May 2008

PS: Ironically AutopanoPro does a great job at auto-detecting CP in the images above. Some manual realignment are needed though.