Zi Jian Yew

24GB is sufficient for inference. The memory heavy portion in the network is the attention layers which takes place only in the coarsest downsampled level. We designed the network to...

Hello. If you use the provided ModelNet configuration you should be able to get the reported results. I remember the performance on Modelnet to be quite stable between training runs....

This is likely related to the issue with Minkowski engine as noted in the other issues, e.g. #1. I do hope to resolve this by removing the Minkowski engine dependency...

I’m unable to check for sure since I’m currently traveling. But if I remember correctly that’s just the ground truth overlap ratio. I don’t think the code makes use of...

Hi, depending on the dataset (3DMatch/ModelNet), you first need to follow download the datasets (instructions in the readme) and place them in the folder.

Hi, I don’t have a ready script as this work is quite a long time ago. I recommend Open3D or vtk for plotting. For smaller features you can try changing...

Hi, I'm using the sparse tensors from Minkowski Engine as a quick way to perform the downsampling. Your issue might be due to a bug in your version of Minkowski...

Just to clarify, the code should work on PyTorch 1.9.1 (which was used in the actual training). I used v1.10 because I had problems using v1.9 for PyTorch3D installation: I...

Fig 5 does not show the attention but visualizes the locations of the keypoints and their corresponding locations. The network outputs all of these quantities so it should be quite...

This is related to #1. Unfortunately I'm not able to replicate the problem on my machine. You might find it useful to install MinkowskiEngine using the commands listed in [this...