Scaling Issue

[chrisgrill]

I am trying to stitch together images that have been collected by pulling a camera over an object of interest. All the images are the same size and only represent translations of the camera.

The issue I am seeing is that the final stitched image has a "tailing off" effect where each image is slightly smaller than the one before it.

The scaling happens somewhere in here:
warped_low_imgs = list(warper.warp_images(low_imgs, cameras, camera_aspect)) warped_low_masks = list(warper.create_and_warp_masks(low_sizes, cameras, camera_aspect)) low_corners, low_sizes = warper.warp_rois(low_sizes, cameras, camera_aspect)

I used the example Jupyter notebook to try to troubleshoot this. Here is the mask:

Any help or suggestions greatly appreciated.

[lukasalexanderweber]

If the camera is only translated you should use affine stitching

[lukasalexanderweber]

maybe first check if your images are really suitable for the stitching package: #264

[chrisgrill]

The images should be suitable for affine stitiching but when using the affine warper it comes out even worse. It still has a trailing effect and the individual images are warped in a way that seems to cause the edges to end up in the center of the image.

[chrisgrill]

It seems like it is the estimated camera parameters. If I substitute the estimated camera rotation matrix with the identity matrix, the images do not get scaled and look normal, but the stitcher does nothing with them. Is that matrix relative to the prior frame?
Here is the mask that I end up with:

[lukasalexanderweber]

If you have a handheld camera and translate along a scene this might lead to your problem, but you can also share your images if you want

[lukasalexanderweber]

Do you stitch all Images at once? If not it might be related to #238

[chrisgrill]

This was caused by the mounting of the camera not being totally perpendicular to the surface being scanned. I thought the camera estimation steps would correct for it but it seems they don't. The solution was to perspective warp the images before trying to stitch them.

img_size = (frame.shape[1], frame.shape[0])
src = np.float32([[0, 0], [640, 0], [640, 480], [0, 480]])
dst = np.float32([[0, 0], [640, -50], [640, 510], [0, 480]])
transform_matrix = cv.getPerspectiveTransform(src, dst)
frame = cv.warpPerspective(frame, transform_matrix, img_size)

[lukasalexanderweber]

Great that it works, thanks for the Feedback! How did you figure out the offset?

[chrisgrill]

I just eyeballed it for now. I will probably create a calibration routine that does some edge detection on two sides of a target and warps them parallel. Now I have some questions on the camera adjuster but I will start another discussion on that. Thanks for the help.