Open syqzdy opened 5 years ago
Thanks for your code! Looking forward to the code of 3dmatch being open sourced as soon as possible! By the way, I'm curious about the results of rotated 3dmatch benchmark if we give xyz input (instead of 4d-ppf feature) to the 3d capsule net. Thanks!
@kxhit By construction, our 3d point capsule network is neither rotation invariant nor resilient. In the case of XYZ input, you will have to heavily augment training set, which might significantly increase training times. Avoiding this and providing a theoretical invariance guarantee is the entire point of PPF-FoldNet. Our capsule nets borrow the same idea for this experimentation. Hopefully we will be releasing some form of the code for this soon.
@tolgabirdal but in your paper, you mentioned that in
in "Do Our Features Also Perform Well Under Rotation?"
Our unsupervised capsule AE, however, is once again the top performer, surpassing the state of the art by ∼ 12% on 2K-point case as shown in Tab. 2. This significant gain justifies that our encoder manages to operate also on the space of 4D PPFs, holding on the theoretical invariances. so ?
what is the radius used on 3Dmatch to extract patches?
As mentioned at the end of the README file, this is planned. We hope to release it in close future.