Fixing the panorama's horizon
Contents
Context
Skyline is a critical element of a panorama. Indeed, human eye is accustomed to the continuity of skylines. When a stitching presents such a default, the overall panorama's quality is affected.
In this tutorial, we will solve this problem on a real world example.
Example 1
The shooting
- Photographer : Leif Roar Strand
- Number of images : 36 photographs
It is important to notice than the shooting has been done under windy weather.
The stitching using Autopano Giga
Using the default settings, Autopano Giga created the following panorama:
Here, we can see 2 kinds of stitching problems:
- Wavy horizon
- Stitching errors
Even if we use the Automatic horizon tool, the skyline issue is not fixed, we will see why.
The control points
First of all, let's have a look to the panorama's structure, to check where the control points have been detected.
The initial structure seems ok at first glance, even if some links have high RMS values.
Moreover, we can notice that the 2 lines are not linked together on the whole panorama length (Autopano didn't found any control points between those images).
Considering the scene (moving clouds and sea), it is not really surprising.
Here are the control points between photos 11 and 12 (RMS : 5.83)
Each of those control points are wrong between the images there! Indeed clouds have moved during the shooting, so the points don't link stationary objects.
Then a parallax issue occurs: the points nearest to the skyline move less than the farest points (in the image plan). Therefore, we cannot consider them as valid.
Let's modify the control points as shown below:
We remove all the control points, on the cloud and on the water, then we search points near the skyline.
This correction must be applied on all the links of the upper line, regardless of the RMS value.
In case an image pair contains control points on the foreground, it requires to remove them and to constraint the detection on the skyline level (taking care to search points doing quite large search ing areas to find control points).
Let's have a look to the resulting panorama (RMS : 2.35):
We can check that the skyline doesn't contain any stiching error.
We can see the clouds are no more perfectly aligned, but it is correct due to the weather conditions.
Now it requires to fix the wavy horizon.
To avoid wavy shape on the skyline, we will add links between the 2 photos rows, eg between images 16 and 34 then images 15 and 33.
The structure of the panorama is now well constrained, with control points showing almost no parallax, we can use the Move panorama tool (Panorama mode) to adjust the horizon.
The final results is:
Conclusion
This example shows that automatic detection doesn't consider the nature of the scene photographied.
Indeed, it is almost impossible to guess if a move is caused by external conditions (the wind here) or the shooting itself.
In such cases, we can use this method above to constrain the control points detection.