BD downsample

[hlx-d]

When I tested the vid4 dataset with pre-trained model ”basicvsr_vimeo90k_bd_20210409-0154dd64.pth”, I got
Eval-PSNR: 20.23476129041623
Eval-SSIM: 0.5117247318824256
I use the following code for BD downsampling：
`
def gaussian_downsample(x, scale=4):

assert scale in [2, 3, 4], 'Scale [{}] is not supported'.format(scale)

def gkern(kernlen=13, nsig=1.6):
    import scipy.ndimage.filters as fi
    inp = np.zeros((kernlen, kernlen))
    # set element at the middle to one, a dirac delta
    inp[kernlen // 2, kernlen // 2] = 1
    # gaussian-smooth the dirac, resulting in a gaussian filter mask
    return fi.gaussian_filter(inp, nsig)

if scale == 2:
    h = gkern(13, 0.8)  # 13 and 0.8 for x2
elif scale == 3:
    h = gkern(13, 1.2)  # 13 and 1.2 for x3
elif scale == 4:
    h = gkern(13, 1.6)  # 13 and 1.6 for x4
else:
    print('Invalid upscaling factor: {} (Must be one of 2, 3, 4)'.format(scale))
    exit(1)

C, T, H, W = x.size()
x = x.contiguous().view(-1, 1, H, W)  # depth convolution (channel-wise convolution)
pad_w, pad_h = 6 + scale * 2, 6 + scale * 2  # 6 is the pad of the gaussian filter
r_h, r_w = 0, 0

if scale == 3:
    r_h = 3 - (H % 3)
    r_w = 3 - (W % 3)

x = F.pad(x, [pad_w, pad_w + r_w, pad_h, pad_h + r_h], 'reflect')
gaussian_filter = torch.from_numpy(gkern(13, 0.4 * scale)).type_as(x).unsqueeze(0).unsqueeze(0)
x = F.conv2d(x, gaussian_filter, stride=scale)

x = x[:, :, 2:-2, 2:-2]
x = x.view(C, T, x.size(2), x.size(3))
return x`

Is the way of downsampling correct?
Can you provide a script for generating the BD down-sampling data of the Vidmeo dataset?
Looking forward to your reply, thank you！

[ckkelvinchan]

I used MATLAB to downsample the images. You can try this script.

[hlx-d]

I used MATLAB to downsample the images. You can try this script.

I used MATLAB to downsample the images. You can try this script.

OK. I'll try it. Thank you！

[ckkelvinchan]

Closing due to inactivity, please reopen if there are any further problems.