ROI returned by cvGetOptimalNewCameraMatrix has wrong width & height

[devernay]

System Information

OpenCV python version: 4.8.0.76
Operating System / Platform: macOS
Python version: 3.10.6

Detailed description

I noticed here that cvGetOptimalNewCameraMatrix returns (0,0,width-1,height-1) when the distortion parameters are zero.

Looking more closely at the code, I see that the internal function icvGetRectangles finds the rectangles by sampling points in the (0,width-1)x(0,hieght-1) range, and then computes the output ROI sizes as (maxX-minX,maxY-minY), which of course misses one pixel in each dimension!

I think the code in icvGetRectangles should add 1 pixel to both dimensions, because the max value should be included.

Steps to reproduce

Test code:

import cv2
import numpy as np
# zero distortion
distortion_params = np.zeros(8)
width=1920
height=1080
ffactor=1.6
# intrinsics with a centered principal point
K1 = np.array([[width*ffactor, 0, (width-1)*0.5],
              [0,width*ffactor, (height-1)*0.5],
              [0,0,1]])
# intrinsics with a non-centered principal point
K2 = np.array([[width*ffactor, 0, (width-1)*0.5+0.05],
              [0,width*ffactor, (height-1)*0.5-0.04],
              [0,0,1]])
for K in [K1,K2]:
    newK, roi = cv2.getOptimalNewCameraMatrix(K, distortion_params, (width, height), 0)
    print(np.linalg.norm(K-newK))
    print(roi)

Output:

4.547473508864641e-13
(0, 0, 1919, 1079)
4.547473508864641e-13
(0, 0, 1919, 1079)

both rois should be (0, 0, 1920, 1080).

Issue submission checklist

• I report the issue, it's not a question
• I checked the problem with documentation, FAQ, open issues, forum.opencv.org, Stack Overflow, etc and have not found any solution
• I updated to the latest OpenCV version and the issue is still there
• There is reproducer code and related data files (videos, images, onnx, etc)

[devernay]

note: previous icvGetRectangles fix, from 4.8.0: #23305

[f-dy]

bug introduced here:

#23305 changed here the pixel sampling from 0 to width-1, which is good, but then the ROI width should be maxX-minX+1, not maxX-minX

[vrabaud]

The error is actually a bit different. Adding +1 is wrong as the rectangle is a Rect_<double> (switching it where you suggest gets the Calib3d_GetOptimalNewCameraMatrixNoDistortion.accuracy test to fail). The problem arises when converting the Rect to ints for the output. I am preparing a PR.

[devernay]

Yes it should do floor(x-0.5) on the minimum bound and ceil(x+0.5) on the second

[devernay]

Or something like that. To take into account the footprint of a pixel

[gchhablani]

Are there any updates on this one?
Would be nice to get this fix.
Thanks!