The S3DIS dataset can be downloaded here (4.8 GB).
Download the file named Stanford3dDataset_v1.2.zip
, and move it to data/Stanford3dDataset_v1.2
. You may also specific your own data directory by changing the path
argument in train.py
.
Compile the C++ extension modules for python located in cpp_wrappers
. Open a terminal in this folder, and run:
sh compile_wrappers.sh
- The code has been tested on one configuration:
- PyTorch 1.8.1, CUDA 10.1
We train the network on a Tesla V100 gpu (to maintain the batch size). It will take a few more time in the first training. The pretrained model can be found here. Simply run:
python train.py
You may reduce the batch_num
in train.py
for some smaller 12GB gpus (train_tiny.py
).
The models are saved in results/train/checkpoints/
every 10 epochs.
To test the model current_chkp
in the previous run:
python test.py --log ./results/train
And to test a model in epoch n:
python test.py --log ./results/train --model epoch_0099.tar
The S3DIS data processing was borrowed from KPConv