Good catch! I think the variable names are wrong, but the output should be correct.

Unfortunately I don't have write privileges and can't update the code :/

Yes, the output is not affected since the inverted meaning is inverted again when accessing the heatmap and for the grid_sample.

When porting this to other languages, e.g. C++, you have to sample the feature map with (x,y) tuples. The python code currently suggests that these should be (y,x) tuples, resulting in the wrong descriptors.

@christian-rauch I try a small example, and I feel no wrong:

import numpy as np
a = np.random.randn(3, 3)
a
Out[4]: 
array([[ 0.24942095,  0.8585088 , -0.1106783 ],
       [ 0.07011566,  0.21370989,  0.81164736],
       [ 2.40070676, -0.11000709, -1.76859847]])
np.where(a > 0)
Out[5]: (array([0, 0, 1, 1, 1, 2]), array([0, 1, 0, 1, 2, 0]))

Where do I understand wrongly?
np.where return (x, y)

Where do I understand wrongly?
np.where return (x, y)

The first row in your matrix a is indeed the first column of the image. That means an index in your first matrix row is the index of the first image column, which is 'y'.

You will see this better when you visualise the data:

import numpy as np
import matplotlib.pyplot as plt

if __name__ == "__main__":
    a = np.zeros(shape=(5,7))

    # set pixel at x=5 and y=2 to 1
    a[2,5] = 1

    # find the coordinate of value '1'
    y,x = np.where(a==1)
    print("point (x,y): ",x[0],",",y[0])
    plt.imshow(a)
    plt.show()

The image is:

And the 1 (yellow square) is located at:

point (x,y):  5 , 2

@christian-rauch Oh I got it, thanks for your patient and vivid explanation.

BTW, could you provide a script that can draw Qualitative Results of a pair of images like paper show(see followiing)? Thank you for your help

@laisimiao I opened pull request #14 to fix this issue. Is this sufficient to demonstrate that it works?

Btw, at

Hi, I think @christian-rauch is right, and should fix, because
samp_pts[0, :] = (samp_pts[0, :] / (float(W)/2.)) - 1.
the samp_pts[0] represent x axis