[bechmark] Add benchmark for all operations

[ternaus]

It would be great to have a benchmark in Readme and / or documentation for kornia-rs vs OpenCV for all similar operations.

In https://dev.to/viglovikov/jpeg2rgb-array-showdown-libjpeg-turbo-vs-kornia-rs-vs-tensorflow-vs-torchvision-2mnh

it was unexpected to see OpenCV being twice slower. Similar numerical comparisons for other operations would be ultra useful in decisions about kornia-rs integrations to other packages.

[edgarriba]

@ternaus thanks for raising interest. I wrote an initial benchmark here but for the resize function only and not well designed nor using the proper tools: https://github.com/kornia/kornia-rs/blob/main/py-kornia/benchmark/resize_benchmark.py

i will create something more stable starting by the minimum functionality we expose in the python api. Besides, would be great if you can help to identify what functions you are interested in that we can port/measure from opencv.

[ternaus]

Thanks!

Currently, in Albumentations we use:

• cv2.inpaint
• cv2.GaussianBlur
• cv2.line
• cv2.cvtColor
• cv2.imread
• cv2.resize
• cv2.addWeighted
• cv2.LUT
• cv2.meanStdDev
• cv2.circle
• cv2.blur
• cv2.imdecode
• cv2.createCLAHE
• cv2.transform
• cv2.calcHist
• cv2.equalizeHist
• cv2.merge
• cv2.multiply
• cv2.subtract
• cv2.flip
• cv2.perspectiveTransform
• cv2.warpAffine
• cv2.getRotationMatrix2D
• cv2.getAffineTransform
• cv2.distanceTransform
• cv2.threshold
• cv2.Canny

These grep shows me about OpenCV, and many other that are done on the numpy level, but, I guess, could, in theory, be done faster with kornia-rs

[edgarriba]

Awesome, give a look at the rust documentation because few of them are already there but not exposed to python yet. I’ll be moving slowly so that we can benchmark each of them. Besides, almost each of them are already implemented in Kornia torch which the strategy is to execute in batch /gpu the augmentations pipeline.

[edgarriba]

And abit of curiosity, why do you use Canny since it’s an expensive edge detection operator. And why multiply, subtract, merge, meanStdDev why can be done directly in numpy, or is the last slow ?

[ternaus]

Strory behind adding Canny I do not remember - I did not even know that it is slow since you told me :)

For numpy vs. cv2, cv2 could be a few times faster for some data types than a similar numpy operation.

For example, in the recent benchmark vs Kornia, torchvision, Augly, ImgAug, we found that we are not the fastest for GaussianNoise.

Checked the code:

def gauss_noise(image: np.ndarray, gauss: np.ndarray) -> np.ndarray:
    image = image.astype("float32")
    return image + gauss

And it does not look cv2 optimized at all. I did not benchmark, but this code could work faster.

def gauss_noise_optimized(image: np.ndarray, gauss: np.ndarray) -> np.ndarray:
    if image.dtype == np.float32:
        gauss = gauss.astype(np.float32)
        noisy_image = cv2.add(image, gauss)
    elif image.dtype == np.uint8:
        gauss = np.clip(gauss, 0, 255).astype(np.uint8)
        noisy_image = cv2.add(image, gauss)
    else:
        raise TypeError("Unsupported image dtype. Expected uint8 or float32.")
    return noisy_image