The pytorch implementation of flownet3d based on WangYueFt/dcp, sshaoshuai/Pointnet2.PyTorch and yanx27/Pointnet_Pointnet2_pytorch
PyTorch>=1.0: https://pytorch.org
scipy>=1.2
numpy
h5py
tqdm
Install this library by running the following command:
cd lib
python setup.py install
cd ../The processed Flyingthings3d data is provided here for download (total size ~11GB).
Then run the following command to train:
python main.py --exp_name=flownet3d --dataset_path=xx/yywhere xx/yy is the dataset path
All of the following experiments were tested on a TITAN RTX
- GPU memory usage:
| batch size | flownet3d(GB) | flownet3d_pytorch(GB) |
|---|---|---|
| 16 | 16.9 | 9.1 |
- Training time per epoch on Flyingthings3d dataset:
| batch size | flownet3d(min) | flownet3d_pytorch(min) |
|---|---|---|
| 16 | 6.7 | 3.4 |