facebookresearch/active_indexing

Official implementation of "Active Image Indexing"

active_indexing's Issues

Sobel Kernels in JND

In file activeindex/attenuations.py, in the JND class, there seems to be a mistake with the sobel kernels. In particular, the kernel for computing x and y gradients are the same:

class JND(nn.Module):
    """ https://ieeexplore.ieee.org/document/7885108 """

    def __init__(self, preprocess = lambda x: x):
        super(JND, self).__init__()
        kernel_x = [[-1., 0., 1.], [-2., 0., 2.], [-1., 0., 1.]]
        kernel_y = [[-1., 0., 1.], [-2., 0., 2.], [-1., 0., 1.]]

They seem to both be kernels for the x-gradient. It also seems that at no point in time is the kernel_y transposed or used differently from kernel_x. Is this a bug or is this expected?

Edit
A quick fix would simply be: kernel_y = [[1., 2., 1.], [0., 0., 0.], [-1., -2., -1.]]