Retrieve corners of the stitched result

[nndei]

Hello,

I am trying to retrieve the corners of a stitched image, as I would like to stretch it out to make the corners coincide with the ones of a rectangle. This is an idea I had in order to minimise the miniaturisation effect observed when stitching together multiple frames in sequence, as displayed here:

What I'd like to do, starting from a simple two-frames stitch, is retrieve the coordinates of the corners highlighted in red, here (note that this is another stitched image, no the simple one I'm working with, but just to understand what are "corners" to me):

I have tried to follow the code and I have drawn the absolute value of final_corners, but the resulting corners are these, which are not what I'm looking for.

Any idea on how to retrieve these?

Thank you and best regards.

[lukasalexanderweber]

I'm quite sure that you can get the points you need from the final_sizes and final_corners values. See here

[lukasalexanderweber]

I don't have that much experience with the corners and sizes. The only thing I can suggest that you closely follow how points and sizes behave at warping and what's there role after blending.

[nndei]

Ok, will do. Thank you

[nndei]

Looking at the documentation a bit deeper, the corners and sizes are obtained by low_corners, low_sizes = warper.warp_rois(low_sizes, cameras, camera_aspect) (cell 44 in the tutorial).
This uses the OpenCV warpRoifunction, which Returns projected image minimum bounding box. (https://docs.opencv.org/4.x/d5/d76/classcv_1_1PyRotationWarper.html#a3afece82103726fd89e9954015b18da3

Therefore, I believe that corners and sizes just represent some sort of bounding box of the warped image. Not the warped image corners themselves. Would make sense because corners are not necessarily on the same horizontal or vertical lines after warping. And achieving them by adding some "size" would only result in a rectangle.

[lukasalexanderweber]

Ok I think I understand. Is there anything else you might want to try?

[nndei]

Honestly, I have no idea how to retrieve them precisely.

Starting with only one image, I've tried to apply the PyRotationWarper.warpPoint method to the four initial corners of image 1. That'd be top_left = (0,0); top_right = (width, 0), and so on.
There's an issue about this method vs the PyRotationWarper.warp too. I am not sure this has been resolved.

Anyway, retrieved top left corners in your tutorial, using weir images, are fine. I mean, the top left corner corresponds to the one you can roughly estimate just by looking at the warped masks in the cropper chapter of the tutorial.
But using my images, they don't really make sense. I obtain negative numbers which means something outside of the image. And projecting them on the warped image doesn't put them in the right place, either.
But yeah, so far I can't retrieve them and I'm stuck with this, but I am working on something else.

I am now trying to make the stitching work using real-time video.

[artes14]

hey if you guys are still interested, I figured out how to get the top-left corners.

minx=min([a for a, b in final_corners])
miny=min([b for a, b in final_corners])
tmp = [(e1-minx, e2-miny) for e1, e2 in final_corners]
for point in tmp:
    panorama=cv.circle(panorama, point, 50, (0,255,0), 10)

it's every x, y coordinates minus the minimum of every x's and y's