Sharp discontinuities in the 360° spherical panorama

[Robert Grübler]

My workflow “poor man's street view” with the Samsung Gear 360 (2017) works quite well, the only annoying thing is the sometimes noticeable breaks in building lines. Like here:

Fig. B01: Inconsistency in the roof line

The template used here to sew the two fisheye images was created in a ‘quick and dirty’ style and therefore seemed to have potential for improvement.

My calibration image is a church interior with a lot of texture throughout the space. In line with my workflow, it is a frame from a time-lapse video:

Fig. B02: Calibration image (full resolution in [1])

Originally, the image had no Exif data, so I copied it over from a photo taken at the same location.

The Hugin project file [2] defines 22 CPs, distributed relatively evenly over 360°. The stitched image has a particularly noticeable discontinuity in the altar steps:

Fig. B03: Equirectangular projection (full resolution in [3])

Fig. B04: Detail from Fig. B03, sharp discontinuity at the steps

Fig. B05: Remapped exposure corrected images as superimposed layers with 50% transparency.

Fig. B05 shows how discontinuity occurs in the step area. The bright vertical stripe in the middle is the overlap area of the two images. The alignment of the images is perfect at the right edge, but unfortunately stitching is not performed there; instead, it is performed more in the middle, where the error is significantly greater.

The harshness of the transition is striking, especially when compared to the Samsung App Gear 360, which shows a similarly large image error but transitions very smoothly:

Fig. B06: Stitching error in the Samsung Gear 360 App. Compare with Fig. B04 Stitching error Hugin

Further attempts to minimize the image errors by

• doubling the control points
• using horizontal and vertical control points
• installing Multiblend

were unfortunately unsuccessful. My research on the Internet was equally unsuccessful. I did find several descriptions of my stitching error—see here, here, and here, but there was no solution.

My gut feeling tells me that something is wrong. But what?

Thanks a lot in advance!

 

 

Anhang

[1] B02f_CalImgAuf3.jpg: Calibration image shot with Samsung Gear 360 dual fisheye lens, f=1.2mm (6mm equ.)

[2] Templ_CalImgAuf3_V3.pto: Hugin project file v2025.0.0 on Windows 10

[3] B03f_Stitched_CalImgAuf3_V3.jpg: Stitched calibration image, spherical panorama

 

[Bruno Postle]

Hi Robert, I would try deleting control points that are not near the likely seam position. This seems counter intuitive, but these pixels will not end up in your final image.

Important, if you do this the optimiser can find all sorts of crazy solutions for the a, b & c lens distortion parameters. To prevent this you need to either: only optimise angle of view, or create lots of vertical control points to properly calibrate the lens distortion (I would create the vertical points).

I would also delete any close control points, ie. The floor might look like a great calibration source, but using it will amplify any parallax problems caused by the slight offset between lenses.

Also, these circular fisheye images are very unlikely to be central in the frame, so you need to optimise d & e parameters separately for each side.

-- 

Bruno

On Sat, 22 Nov 2025, 22:37 Robert Grübler wrote:

My workflow “poor man's street view” with the Samsung Gear 360 (2017) works quite well, the only annoying thing is the sometimes noticeable breaks in building lines. Like here:

My calibration image is a church interior with a lot of texture throughout the space. In line with my workflow, it is a frame from a time-lapse video:

[Robert Grübler]

Hi Bruno, thank you very much for looking at my data.

 

I would try deleting control points that are not near the likely seam  position. This seems counter intuitive, but these pixels will not end up in your final image.

 

That sounds logical to me, even if I don't understand your reasoning. In my words, I would have said that it's because the fish-eye distortion is extreme at the edges, making reprojection very unreliable. Or am I missing something?

 

... or create lots of vertical control points to properly calibrate the lens distortion (I would create the vertical points).

 

Should the vertical control points only be located in the overlap area or also outside it?

 

I would also delete any close control points, ie. The floor might look  like a great calibration source, but using it will amplify any parallax  problems caused by the slight offset between lenses.

 

Good point, thank you.

 

Also, these circular fisheye images are very unlikely to be central in  the frame, so you need to optimise d & e parameters separately for  each side.

 

What do you mean by "separately for each side"?

 

Thank you very much for your valuable advice.

 

Robert

[Bruno Postle]

On Sun, 23 Nov 2025, 20:17 Robert Grübler wrote:

Hi Bruno, thank you very much for looking at my data.

I would try deleting control points that are not near the likely seam  position. This seems counter intuitive, but these pixels will not end up in your final image.

That sounds logical to me, even if I don't understand your reasoning. In my words, I would have said that it's because the fish-eye distortion is extreme at the edges, making reprojection very unreliable. Or am I missing something?

Yes in part. Because your photos overlap, the extreme edges of the circle are not going to be used, so putting control points here isn't going to help and may make things worse.

... or create lots of vertical control points to properly calibrate the lens distortion (I would create the vertical points).

Should the vertical control points only be located in the overlap area or also outside it?

They should be all over the scene, especially in areas that are not overlapped.

I would also delete any close control points, ie. The floor might look  like a great calibration source, but using it will amplify any parallax  problems caused by the slight offset between lenses.

Good point, thank youu 

Also, these circular fisheye images are very unlikely to be central in  the frame, so you need to optimise d & e parameters separately for  each side.

What do you mean by "separately for each side"?

This is advanced usage, only do this if you can't get good results. I assume you are cropping out each circle and saving them as separate files. Even if you crop carefully these are different lenses, so they will be aligned differently relative to the sensors. The d & e parameters tell Hugin where the centre of the lens is relative to the centre of the sensor. So this technique involves you optimising each of the two circles as a different lens altogether.

(Note there is a further more advanced technique where you feed the image with two circles into the hugin project twice, this way you can do all the cropping in Hugin and use the file directly out of the camera, but this can be tricky to set up)

-- 

Bruno

[Robert Grübler]

Bruno, thank you so much!

I will follow your suggestions and report back on the results.

Robert

 

Von: hugi...@googlegroups.com [mailto:hugi...@googlegroups.com] Im Auftrag von Bruno Postle
Gesendet: Sonntag, 23. November 2025 22:31
An: hugin and other free panoramic software <hugi...@googlegroups.com>
Betreff: Re: [hugin-ptx] Sharp discontinuities in the 360° spherical panorama

 

On Sun, 23 Nov 2025, 20:17 Robert Grübler wrote:

--
A list of frequently asked questions is available at: http://wiki.panotools.org/Hugin_FAQ
---
You received this message because you are subscribed to the Google Groups "hugin and other free panoramic software" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hugin-ptx+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/hugin-ptx/CAJV99Zj7hSmWscdFWM5gdQhKCxasL7CsOyUL06FPdNmXO6i-Rw%40mail.gmail.com.

[Robert Grübler]

Bingo – a perfect stitch

 

 

Two ticks and the error is gone.

Thanks Bruno

 

Robert