Interpolation aritifact on pitch rotation

[ivantishchenko]

I'm trying to compare the outputs of equilib to another library https://github.com/BlueHorn07/pyEquirectRotate/

pyEquirectRotate:

# (yaw, pitch, roll) in degree
equirectRot1 = EquirectRotate(h, w, (-90, 3, 0))
rotated_image = equirectRot1.rotate(img_ldr)

equilib :

rots = {'roll': np.radians(0), 'pitch': np.radians(3), 'yaw': np.radians(90)}
equi_img = np.transpose(img_ldr, (2, 0, 1))
rotated_image = equi2equi(src=equi_img, rots=rots, mode='bilinear')
rotated_image = np.transpose(rotated_image, (1, 2, 0))

Both libraries produce the same output visually, but there is an artifact in case of equilib. There is a patch of blue sky right in the middle of the gray floor. Is it possible to get rid of this artifact? I tried all supported interpolation modes: nearest, bilinear and bicubic, but couldn't get rid of the tiny blue patch.

[haruishi43]

@ivantishchenko Thanks for trying our the software! I have not noticed these artifacts during my tests, but I haven't figured out the smartest approach for testing yet (and it's very outdated). I'll check this myself when I have time. Is the image publicly available?
You also seem to be using HDRI, and I haven't tried HDRI inputs for equilib. I'm not sure that HDRI has anything to do with the artifacts, but I'm not really sure yet.

[ivantishchenko]

@haruishi43

Thank you for your quick reply!

It's a publicly available HDRI in OpenEXR format:
https://dl.polyhaven.org/file/ph-assets/HDRIs/exr/1k/limpopo_golf_course_1k.exr

To read into a numpy array HxWxC :

pip install openexr_numpy

from openexr_numpy import imread
img = imread(HDRI_PATH)

[haruishi43]

I tried it with pytorch and the artifacts doesn't seem that severe as numpy:

It's probably a bug in either grid_sample or convert_grid. I'm sorry, but it might take a little bit more time to debug this. My guess is that the artifacts are caused by sampling from the sky region (top row of the image) because of bad sampling or overflow during the grid conversion.