About the training of the sampling location offsets
sstzal opened this issue · 2 comments
Hi, Keunhong
This nerfies is an excellent work! And I have a little question want to ask: When I train the model, the translation in the code as the following img are always very high or low (nearly 0). Have you ever encountered this kind of problem？ How do you limit the range of the translation?
Looking forward to your reply! Thanks!
It is a good question!
This translation filed is referenced in D-nerf
Yuo can have a look at this D-nerf code.
Hi, the weights of the last layer in the deformation MLP are initialized close to the identity, since you want it to stay close to the identity unless it needs to deform. I didn't see any issues with this, but if you want to restrict the deformation to a certain range you could always apply an activation (like a sigmoid of tanh) and scale that however you want.