[Feature Request] Differentiable torch.histc and differentiable equalize

🚀 Feature

Our current implementation used a non-differentiable implementation of equalization. Mainly, it is caused by the non-differentiable torch.histc. We may have different differentiable estimators to estimate the image histogram.

Presumably, I would expect an API as:

def equalize(image, estimator=None):
    if estimator is None:
        hist = torch.histc(...)
    elif estimator == 'gaussian':
        hist = kornia.gaussian_hist(...)
    ...

Additional context

There are some materials already to estimate the image histogram as here.

[shijianjian]

I’ve implemented differentiable image histogram using triangular kernel density estimation using bandwidth = (max - min) / n_bins. This histogram is approximate. How would you suggest to do test_uniform_dist or other tests for approximation?

Is it an implementation from a standard library? We may test the results against other libraries if appropriate.

Or, you may simply test using a manual seed. Using Monte Carlo is even better but I have no idea how to do that. XD.

[shijianjian]

@connorlee77 making sense. Can you file a ticket?