daytan46's profile - activity

Your application is quite specific, I would first code a prototype using Matlab or Python, in order to define an algorithm, and only after that, I would develop the algorithm to work with the Raspberry Pi.

As first draft, I would propose the following steps:

• Database population (using pictures to which you manually associate name of the bird and so on). It could also integrate specific parts of the birds.
• Image alignment (the camera is static, but the bird could be in different position)
• Features definition (shape, colors, tail, beak, etc...)
• Feature collection (one vector for each type of features for example)
• Weight association (some features are more significant than others)
• Definition of a function that estimates the similarity between the features just extracted and the ones in the DB (example: using cosine similarity function)
• Classifier function (which decides the type of bird based on the distances calculated in the previous step)

When you compiled, did you specify

-DBUILD_EXAMPLES=ON

?

That should be enough to find all samples built. Note that in Mac OS X the folder is opencv-2.4.9/samples/, but in Windows it could be in the "release" folder or "bin" folder - just look for the executable you are looking for by using the Windows explorer. If you want to modify the source and recompile it, there should be a file called "link.txt" per each sample, that contains the command to execute to get the new final executables.

I have wide angle lens with very low distortion and a tripod.

I would like to create a spherical panorama (360 degrees on the horizontal axis and 180 degrees on the vertical) by using the Stitching Detailed sample (OpenCV 2.4.9): (https://github.com/Itseez/opencv/blob/master/samples/cpp/stitching_detailed.cpp).

I have 3 questions. Here they are:

1) Pictures are always adjacent. Therefore, if I have

1. first.jpg
2. second.jpg
3. third.jpg

I call the executable by typing:

stitching_detailed first.jpg second.jpg third.jpg

The overlapping region between first.jpg and second.jpg is the left border of first.jpg and the right border of second.jpg; the overlapping region between second.jpg and third.jpg is the left border of second.jpg and the right border of third.jpg and so on...

How can I improve the stitching_detailed code both from the "time needed" and "quality" point of view by using this knowledge?

2) I plan to stitch first vertically and then horizontally. When stitching vertically, I use the "--warp plane" option and then, once I removed the black part, I stitch the respective results horizontally with the "--warp spherical" option. Does anybody have suggestions about how improving this approach (or additional options)?

3) When taking pictures from a corner of a room, for example, the stitching becomes more difficult, because of course the FOV is more limited when the the camera is faced to the wall, and the algorithm doesn't work anymore... Did anybody encountered this issue?

I realize that all three questions are quite generic, but my hope is that, if answered, they will help many other people who are trying to use this useful OpenCV class at its best!

Thanks.