Using Python3.8, PyTorch 1.6 and Cuda 10.2, as specified in the repository. While running the sample registration using pretrained weights provided in the link attached to the repository I encountered the following error :

`Traceback (most recent call last):
  File "inference.py", line 72, in <module>
    rma_net.load_state_dict(torch.load(args.weight),True)
  File "/home/gridraster/anaconda3/envs/rmanet/lib/python3.8/site-packages/torch/serialization.py", line 585, in load
    return _legacy_load(opened_file, map_location, pickle_module, **pickle_load_args)
  File "/home/gridraster/anaconda3/envs/rmanet/lib/python3.8/site-packages/torch/serialization.py", line 755, in _legacy_load
    magic_number = pickle_module.load(f, **pickle_load_args)
_pickle.UnpicklingError: A load persistent id instruction was encountered,
but no persistent_load function was specified.

Can you help me fix this issue? What is the pickle version used in the repository while pickling the pretrained weights?

Thanks for your excellent work!

I had some problems when I tried to test RMA-Net on my own data (synthetic data generated from the SMPL model). I have checked my input data (source & target) were stay in a similar coordinate system but different size as your sample data. My result didn't look as good as you show. I have tried to use your sample data, and the results were all perfect. How should I do to get a better result? or if your method has some limitations (maybe it is not suitable for arbitrary body registration)? (The red point cloud is the source, the blue one is the target and the green one is the result from RMA-Net) (The yellow one is my data and another one is your sample data. I am not sure if they are in the same coordinate system but they look similar. )

Thanks for your excellent work and nice code!

I have downloaded the code and checkpoint, and tested it one the samples you gave. I got very good results, then I hoped to test my own files.

I used the SMPL model to generate two poses of the same 'person' (they were set in the same coordinate system), and tested it with the model you trained, but the results were very strange. What prerequisites does this method have for the input data? Should I do any preprocess before deformation?

Hi,

Are there plans to have the pre-trained models available outside Baidu. It looks like currently it requires creating an account on the system in order to download the models. Creating the account though is not straightforward.

Thank you,

When testing on a custom dataset I am getting a deformed result. The source should ideally align itself to the target and take the shape of the target too right?

My data consists of a simulated point cloud of an object (as source) and a point cloud generated from a real image of the same object (as target). This results in a deformed output that is also not aligned perfectly. Is there any explanation for this? Can you suggest a modification to solve the issue?