Questions about the function process_high_res_im about cascadepsp

[hkchengrex]

It is to speed up processing.
If the object is really small it probably cannot be caught in the global step anyway. That being said you can still tune these thresholds for better results.

[GOZGOXGOYGO]

I'm not sure if here is a little mistake. It seems that the choice value should be of 0, 1, 2 or 3. Thanks.

def get_random_structure(size):
    choice = np.random.randint(4)

    if choice == 1:
        return cv2.getStructuringElement(cv2.MORPH_RECT, (size, size))
    elif choice == 2:
        return cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (size, size))
    elif choice == 3:
        return cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (size, size//2))
    elif choice == 4:
        return cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (size//2, size))

[hkchengrex]

Ah, it is a silly mistake. Thank you for pointing it out! 👍
Luckily it should not cause significant issues.
We will just leave it there and add a comment since our model is already trained with this bug.

[hkchengrex]

On second thought I will fix it and add a comment saying that the provided model is trained with the mentioned bug.
I have no idea how this silly bug can sneak into the code...

[GOZGOXGOYGO]

Thanks! It doesn't matter. This is undoubtedly an outstanding work. Fixing this little bug would theoretically yield better performance.