How T-Net is initialized as identity matrix?

[adosar]

According to the https://arxiv.org/abs/1612.00593 paper, the T-Net is initialized as identity matrix.

I fail to understand how this is the case when looking into the STNkd code.

If I understand correctly:

        iden = Variable(torch.from_numpy(np.eye(self.k).flatten().astype(np.float32))).view(1,self.k*self.k).repeat(batchsize,1)
        if x.is_cuda:
            iden = iden.cuda()
        x = x + iden
        x = x.view(-1, self.k, self.k)
        return x

This snippet adds the identity matrix to the regressed matrix. How the identity initialization is achieved?

[purple24769]

🤔I feel that the matrices obtained through the STNkd network have relatively small values. Therefore, adding the identity matrix does not lead to significant multiplication effects with the original point cloud data. And this stabilizes the training process by avoiding overly aggressive transformations on the input data during the initial stages of training.