Adversarial Scene Flow

Self-Supervised Scene Flow estimation via Adversarial Metric Learning.

Get started

To use this repository first clone this repository in your system.

conda create --name aml python=3.6
conda activate aml
git clone https://github.com/VictorZuanazzi/AdversarialSceneFlow.git
cd AdversarialSceneFlow

Downloading data:

Install pytorch_geometric to use the ShapeNet pointclouds.

Download FlyingThings3D, KITTI without ground and KITTI with ground

Download and install Lyft Level 5

pip install lyft-dataset-sdk

Models for flow extraction:

We encorage you reading the documentations before installing the external libraries.

Clone and install Kaolin :

(From experience, Kaolin does not integrate greatly with different GPUs, execute the following commands using the GPU model you will use in your experiments).

git clone https://github.com/NVIDIAGameWorks/kaolin.git
cd kaolin
python setup.py install
cd ..

Clone and install Pytorch's implementation of FlowNet3D (requires GPU to run):

cd AdversarialSceneFlow
git clone https://github.com/hyangwinter/flownet3d_pytorch
cd flownet3d_pytorch/lib
python setup.py install
cd ../../

Clone and install PointPWC-net:

git clone https://github.com/DylanWusee/PointPWC.git
cd PointPWC/pointnet2
python setup.py install
cd ../../

Training

python main.py --exp_name aml --train_type triplet --flow_extractor flownet3d --dataset shapenet

Evaluate

python main.py --exp_name aml_test --train_type evaluate --flow_extractor flownet3d --dataset shapenet --load_model <path/to/model/best_model.pt>