Use much more time vs opencv and python image.resize().

[zhaocc1106]

I resize one image to 448 * 448:

pillow-resize use 32382 us.
opencv resize use 337 us.
python image.resize() use 7209 us.

codes like:

// pillow resize
cv::Mat thumbnail_img = PillowResize::resize(image, cv::Size(image_size, image_size),
                                                 PillowResize::InterpolationMethods::INTERPOLATION_BICUBIC);

// opencv resize
cv::Mat thumbnail_img;
    cv::resize(image, thumbnail_img, cv::Size(image_size, image_size), cv::INTER_CUBIC);

// py image resize
thumbnail_img = image.resize((image_size, image_size))

[zhaocc1106]

Any updates?

[marcov868]

I made the tests with the image you shared and I got quite different results:

• pillow-resize: 26448 μs
• OpenCV (4.10.0): 2755 μs
• PIL (python 3.12 + PIL 11.0.0): 29189 μs

In my environment, the results between pillow-resize and PIL seem comparable, which makes sense since they are based basically on the same code. OpenCV is way faster for me too, but for "only" one order of magnitude. I never dug too much inside the OpenCV code, so I'm not sure the reason behind this difference. I guess it is because of a more optimized code and probably due to the non-exact antialiasing algorithm.

Back to your case, I can't explain your results since they are quite different from mine. Can you share more details about the environment and the compilation methods of the scripts?