search

yewzijian / 3DFeatNet

3DFeat-Net: Weakly Supervised Local 3D Features for Point Cloud Registration
MIT License
220 stars 46 forks source link

How to train the network #1

Closed fukexue closed 5 years ago

fukexue commented 5 years ago

Hello, I'm sorry to interrupt. Is it convenient to provide a training version of the code?

yewzijian commented 5 years ago

Hi, the training code is available in the repository. We provide a bash script train.sh to perform the training on our data, which internally calls train.py. Please see the main README.md for details.

fukexue commented 5 years ago

Thank you for your reply.

fukexue commented 5 years ago

Sorry to bother you, I only generated train folder and train. TXT using two script files of your pre-processing Oxford dataset, but train. py needs cluster folder. How does cluster folder generate?

fukexue commented 5 years ago

Tips for errors as shown in the figure during training. 2019-01-20 13-31-42

yewzijian commented 5 years ago

Hi @fukexue, the cluster folder is the test data used to used to generate Table 2 in our paper. See the README.md in the scripts_data_processing folder for the download link. You will likely get better results than the values reported in our paper as we've tweaked the training protocol slightly.

Alternatively you can just comment out the parts which use these clusters; they do not affect the training in any way.