Question regarding neighbor searching. about shift-net_pytorch HOT 4 CLOSED

Hi, it is not quite easy to explain unless you debug here: https://github.com/Zhaoyi-Yan/Shift-Net_pytorch/blob/master/models/shift_net/InnerShiftTripleFunction.py#L36,

        cosine, latter_windows, i_2, i_3, i_1 = bNonparm.cosine_similarity(former_all.clone(), latter_all.clone(), 1, stride, flag)

        _, indexes = torch.max(cosine, dim=2)

        mask_indexes = (flag==1).nonzero()[:, 1].view(ctx.bz, -1)

        non_mask_indexes = (flag==0).nonzero()[:, 1].view(ctx.bz, -1).gather(1, indexes)

        idx_b = torch.arange(ctx.bz).long().unsqueeze(1).expand(ctx.bz, mask_indexes.size(1))
        # set the elemnets of indexed by [mask_indexes, non_mask_indexes] to 1.
        # It is a batch version
        ctx.ind_lst[(idx_b, mask_indexes, non_mask_indexes)] = 1

_, indexes = torch.max(cosine, dim=2) is what you want to find and ctx.ind_lst is the matrix P.

from shift-net_pytorch.

Hey @Zhaoyi-Yan,

I hope you are doing well. I lost your contact. Would be great to connect.
I was wondering if you were interested in pursing inpainting but in different data format.

I am currently one the main contributor of TorchPoints3d: https://github.com/nicolas-chaulet/torch-points3d.

And I guess ShiftNet would be pretty interesting to perform in-painting / densification in 3d.

Best regards,
T.C

from shift-net_pytorch.

@Zhaoyi-Yan, thank you for your answer, it is really not easy and as I understand you're calculating Cosine Similarity for a pixel from missing region and then shift the most suitable pixel from know region in that place.

from shift-net_pytorch.

@tchaton
I am sorry that I have devoted myself to another area, so I am afraid I do not have time to performing inpainting in 3d.

from shift-net_pytorch.