search

FabianFuchsML / se3-transformer-public

code for the SE3 Transformers paper: https://arxiv.org/abs/2006.10503
475 stars 69 forks source link

About the experiment on ScanObjectNN #2

Open DLwbm123 opened 3 years ago

DLwbm123 commented 3 years ago

Hi! Great work! Could you please release the SE(3)-Transformer code on ScanObjectNN? Thanks!

FabianFuchsML commented 3 years ago

Hi! Great to hear that you are interested! Unfortunately, the IP situation specifically around that part of the code is quite challenging. I am on it and I'll post something if/when there are updates.

linqingfan commented 2 years ago

Is it possible to describe in pseudo code how to process point clouds (RGBD) with fibres?

doloresgarcia commented 1 year ago

Hi Fabian and authors! I think I am running into the same issue as I am trying to apply the se3 transformer to my dataset, and I have a dgl graph with variable number of edges (sometimes fully connected or knn). How does this come together with the fiber data structure?

FabianFuchsML commented 1 year ago

Conceptually, the important bit about the fiber structure is that you have different ‘types’ of features (scalars, 3D vectors, etc.) attached to the same node. So, in principle, this is orthogonal to the question of how many edges (or nodes, or edges per node, …) you have. I would think that dgl will handle this for you, but I don’t quite remember the details of dgl. General piece of advice: the Nvidia implementation of the SE(3)-Transformer is much much faster (orders of magnitude). So, nowadays I’d advise working with that one (they also use dgl):

Nvidia version here

JosefT-TuDortmund commented 1 year ago

Hi Fabian! I just want to ask whether you can provide any update regarding the availability of the ScanObjectNN code?

FabianFuchsML commented 11 months ago

Hi Josef!

Thank you for reaching out and showing interest in the SE3 Transformer.

At present, I'm quite occupied and unfortunately don't have the bandwidth to provide updates to this repository.

However, I find the Nvidia implementation of the SE3 Transformer I linked above quite exciting. It's a newer and much more efficient implementation. I am sure it can be adapted straight-forwardly to work with ScanObjectNN.