[biz-engineer]

Image stitching may be done for the purpose of creating super resolution panoramas. So we may be tempted to use our best bigger sensor camera, but the bigger camera isn't the best choice for image stitching and there is reason for that: how parallax error prevents stitching software to achieve seamless image stitching.

Experience shows that wide angle image stitching without lens rotation around the nodal point, with loads of parallax error can still be stitched without problem and the way it can be achieved is by decreasing the step angle amount between each frame. What that means is that taking more pictures with a smaller sensor camera (e.g mobile phone or compact camera or micro4/3 or apsc) is better than taking fewer pictures with a larger sensor camera (e.g full frame or medium format camera). The ideal camera for robust image stitching even with parallax error would be a camera with an large expect ratio sensor (e.g 10:1 ratio) panned horizontally in portrait orientation such that the step angle between frames is very small e.g 5 degrees rotation is much better than 30 degree rotation with 50% overlap between frames. An experiment showed that wide angle image stitching in presence of parallax error produced seamless panoramas when the images are "thin sliced" by ~ 5 degrees steps before the stitching process.

So, here are my findings for seamless image stitches even in case of parallax error:
- The camera should be in portrait orientation for an horizontal panorama, so that parallax error amount is minimized between frames
- The camera should be in landscape orientation for vertical panos, so that parallax error amount is minimized between frames
- Moving the camera horizontally in landscape orientation is the worst condition for image stitching software in presence of parallax errors
- If the pano is a square image mosaic, use square crop mode and zoom in further to reduce the step angle between frames
- For seamless image stitching, a camera with the smaller sensor is better than a camera with large sensor

This is counter intuitive because you would think of using a larger sensor camera if you want more resolution by means of stitching multiple image frames.

---

Last edited by biz-engineer; 01-14-2022 at 03:57 PM.

[photoptimist]

Parallax errors scale with D*2*sin(Theta/2) where D is the distance from the sensor to the lens' nodal point and Theta is the swing angle.

You are right that taking fewer shots with a larger-format, higher-resolution camera will risk worse parallax errors than taking more shots with a smaller-format, higher-resolution camera. In fact the larger camera is doubly-bad both due to the larger swing angle and due to the larger nodal distance typically found with a larger camera with its larger lens.

However....

1. Nothing stops the pano shooter from taking lots of small steps with a larger-format camera -- digital images are almost free.

2. Nothing stops the pano shooter from mounting their larger-format camera such that the pivot point is quite close to the nodal point.

3. Smaller-format, lower resolution cameras suffer from requiring multiple rows of shots (a 100 MPix medium format camera in the recommended portrait orientation collects a row 11500 pixels tall while a 20 MPix M4/3 camera collects a row only 5100 pixels tall so it requires 3 overlapping rows to get as much resolution). Collecting multiple rows increases the risks of subject motion or changes in lighting during the collection process.


All that said, your analysis does show the merits of narrow-angle stepping rather than trying to get away with fewer shots and bigger angle steps. And it shows that getting good results with image stitching requires some thought.

[Michail_P]

I appreciate your thoughts on stitching, as it is not a common technique for me and would certainly have to try that. Panoramic views are offered generously so I’ll get my tripod and have some fun. Are there any suggestions for stitching software? I use Ps or darktable most of the times, but I think there must be something better for that specific task.

[biz-engineer]

1. Nothing stops the pano shooter from taking lots of small steps with a larger-format camera -- digital images are almost free.

That's right. While I was testing my pano head setting at home, I shot the dinning room table with the chairs around, very close to the camera, so it was very challenging for Hugin to line-up everything. Then I shot the same scene again with the camera portrait oriented, and rotating the camera 1/10 of the frame for every exposure, I try to stitch again, Hugin still unable to line-up every frame perfectly. Then I simply batch cropped all images in the middle vertically, leaving about 20% of the image width but keeping 100% of the image length, fed Hugin with the batch output.... and I was shocked, Hugin lined-up every frame perfectly, and the total image was super wide angle.

2. Nothing stops the pano shooter from mounting their larger-format camera such that the pivot point is quite close to the nodal point.

Exactly, that's what I was trying to do, but still when the subject matter was 3 dimensional and very close to the camera (closest element of table + chairs was a foot or two feet away from the front of the lens), the stitch output didn't have all lines match, some lines couldn't match. Here, I was trying to test the scenario of a wide angle landscape panorama with a predominant subject in the foreground.

Collecting multiple rows increases the risks of subject motion or changes in lighting during the collection process.

Yes, you are correct. So, we may decide to take a greater or lesser number of exposures, depending on proximity of the foreground elements, if there are moving elements in the scene or if the light is changing fast. If things are moving in the scene we can also intentionally do long exposures to smooth out the junctions. Hugin also allow the masking, it's possible to include or exclude parts of images that moved frame to frame. The thing is, fleeting light is in contradiction with doing long exposures, so in that case it is better to take the least number of exposures and avoid a foreground subject too close to the camera.

---------- Post added 15-01-22 at 08:39 ----------

The No Parallax point measurement challenge:

In order to reduce stitching errors, I've tried to measure the position of the no parallax point for the main focal lengths of the lenses I use for panorama stitching, and write those arca rail positions on a small look up card that I keep in my camera bag for use in the field. I was never able to find the no parallax point very precisely.

I tried four measurement methods:

1) The first method I used was to take two vertical matching lines , one line close to the camera and the other line far from the camera, rotate the camera to see if the near and far lines stay matched, more the camera/lens of the arca rail to reduce the mismatch between lines, check again, fiddle with the camera position to find the position that creates the least amount of parallax, visually. There are two problems with that method: have enough depth of field to see the front and back lines both sharp and at the same time not have too much aperture diffraction such the vertical lines appear blurred at 100% magnification.

2) The second method I've used was to take one distant vertical line and an horizontal tape measure closer to camera, note the tape measure number where the vertical line crosses the tape measure when the line is in the center of the frame. Then rotate the camera to the far right or far left, so that the vertical line is at the edge of the frame. Move the camera/lens are one end of the rail, note the arca rail position number and note the tape measure number when the vertical line crosses. Move the camera at the other end of the rail, vertical line still at the far right or far left of the frame, note the arca rail position and note the tape measure number when the vertical line crosses the tape measure. This gives 4 numbers, that describe a line in an X,Y chart, the zero crossing of the line indicates the theoretical no parallax point of the camera/lens in terms of the arca rail position. I thought that method , in theory, should be more precise than the method 1 (above), but it wasn't.

3) The third method was the quicker variant of method 1, and using a vertical line and a tape measure. Lineup the vertical line in the center of the frame, read the number on the tape measure where the vertical line crosses the tape measure. Rotate the camera so that the vertical line appear at the edge of the frame, note that the number on the tape measure has changed. Without rotating the camera, move the camera/lens position forward or backward with the arca swiss rail until the vertical line matches the measure tape number noted when the vertical line was at a the center of the frame. That method works as well as method 1, but it is quicker.

4) For the fourth method, I took a series of panorama exposures for given camera position on the arca rail. Then I fed the images in Hugin, aligned the images and noted the amount of matching errors reported by Hugin. Loaded another set of images taken with the camera at a different position, aligned images in Hugin and noted the amount of alignment error. So, in theory I should find the camera position on the rail that give the least amount of alignment error in Hugin. But that method never worked for me. The reason is that Hugin find a different cloud of control points for every new set of images, so in the end the amount of alignment error given by Hugin wasn't a reliable quantification of parallax error.

[StiffLegged]

… An experiment showed that wide angle image stitching in presence of parallax error produced seamless panoramas when the images are "thin sliced" by ~ 5 degrees steps before the stitching process…..

That’s an interesting thought biz, thanks!

[AggieDad]

Absolutely have the camera in portrait attitude for horizontal panos.

From my own experience, I find using a longer lens gives me my best panos. I usually use a 100mm or 85mm lens. But a longer lens will result in more images. An 85mm lens will require more than twice as many images as a 35mm lens. A 30-35% image overlap gives me plenty of latitude for stitching.

To keep a pano from looking like a ribbon, take multiple rows of images. Depending on the pano's width, 2 to 4 rows gives you a good wide photo with nice height. Using a longer lens gives nice detail.

[biz-engineer]

Absolutely have the camera in portrait attitude for horizontal panos.

Yes, that's intuitive for many people.

From my own experience, I find using a longer lens gives me my best panos. I usually use a 100mm or 85mm lens. But a longer lens will result in more images

Exactly, and the reason why longer tele lenses are better for stitching panos is because the angle of view is smaller, 500 mm lens has an angle of view of about 5 degrees, that's why when stitching with such lenses rotating around the no parallax point is not necessary.

[AggieDad]

Yes, that's intuitive for many people.


Exactly, and the reason why longer tele lenses are better for stitching panos is because the angle of view is smaller, 500 mm lens has an angle of view of about 5 degrees, that's why when stitching with such lenses rotating around the no parallax point is not necessary.

We're together on this.

[Bob 256]

One of the first panoramas I did using the Hugin pano editor and three hand-held shots using my K-1, amazed me with the stitching job it did. The three shots had a lot of overlap but not that good a registration due to lens distortion and small amounts of varying tilt between images. Nonetheless, I can't see any stitching errors in the final result which I printed as an 8" x 24" panorama. I most certainly didn't align the nodal points of the three shots, nor use a rotating head in this case (I didn't plan to do a panorama at the time). I think the effectiveness of the stitching will depend somewhat on subject matter, the software used, and the quality of the images (parallax reduction, allowances for overlap, and uniform exposure) but using a higher resolution sensor in my case certainly didn't seem to affect the outcome under the poorest of conditions in this case.

I'm not recommending my technique and if I plan to shoot a panorama, I always use a tripod and allow for about 30% overlap, but I haven't seen any issues I would attribute to too high a resolution in the image (other than the fact that the software does have to crank for a really long time for the result).

[Bad Boy]

"- For seamless image stitching, a camera with the smaller sensor is better than a camera with large sensor"

Would not the angle of view be significant rather than the camera sensor size? A larger format camera with longer lens, same angle of view yield better results?

[AggieDad]

"- For seamless image stitching, a camera with the smaller sensor is better than a camera with large sensor"

Would not the angle of view be significant rather than the camera sensor size? A larger format camera with longer lens, same angle of view yield better results?

The field of view is dependent on 2 factors, lens focal length and sensor size. It can be calculated accordingly:

Field (angle) of View (in degrees) = 2 ArcTan( sensor width / (2 X focal length)) * (180/π)

Here is a little table I made for myself when I was first building my automatic pano head. It has values for the APS-C sensor, but simply multiply by 1.5 to get FF values.

[biz-engineer]

Would not the angle of view be significant rather than the camera sensor size?

yes absolutely, but my point is if small angle of view is used, it's exactly what smaller sensors are good at, thank to the crop factor.

[Wheatfield]

I just shoot a bunch of pictures, feed them into Photoshop and let the software sort it out.
It seems to work.