search

Lilac-Lee / PointNetLK_Revisited

Implementation for our CVPR 2021 oral paper "PointNetLK Revisited".
MIT License
42 stars 4 forks source link

Extension of input dim #13

Open StrivedTye opened 8 months ago

StrivedTye commented 8 months ago

Hi Authors,

Thanks for your great work! During training, I want to utilize other geometric features (e.g. normal) except the coordinates (x, y, z). This is, the input shape will become (B, N, 6) instead of (B, N, 3). After doing this extension, there is some problems about the calculation of the jacobi matrix. I still only calculate gradient with respect to x, y, and z: A1 = self.mlp1[0].weight[:, :3, :]. Does it make sense?

Thanks!

Lilac-Lee commented 8 months ago

Hi @StrivedTye, you could consider the ReLU-MLP as ReLU(Ax+b) (see eq. 8). Therefore, you are computing the partial derivative of the network of the input. If your input has size 6, you will get a gradient of size 6, and each dimension is in correspondence with your input. So, your understanding should be correct.

Cheers.