Home

Awesome

Regularizing Self-supervised 3D Scene Flows with Surface Awareness and Cyclic Consistency

<p align="center"> <img src='docs/overview-merge.png'/> <p> <!-- # Results on StereoKITTI dataset --> <!-- ![alt text](docs/overview.png) -->

Installation

<!-- - Install [PyTorch Scatter](https://github.com/rusty1s/pytorch_scatter/tree/master) with CUDA support. -->

DATA

export BASE_PATH='path_where_to_store_data'
tar -xvf data_sceneflow.tgz $BASE_PATH/data/sceneflow

Run Experiments

To run the method on all datasets with final metrics printed on cuda:0, just type:

for i in {0..6}; do python evaluate_flow.py $i; done

where the argument sets the specific datasets according to the following table:

DatasetArgument NumberModel
KITTI t0Neural Prior
StereoKITTI1Neural Prior
KITTI t2SCOOP
StereoKITTI3SCOOP
Argoverse4Neural Prior
Nuscenes5Neural Prior
Waymo6Neural Prior

Experimental results on LiDAR Datasets

<!-- <img style="float: ;" src="docs/table_lidar.png"> --> <p align="center"> <img src="docs/table_lidar.png" /> </p>

Experimental results on StereoKITTI Dataset

<p align="center"> <img src="docs/table_kitti.png" /> </p>

Qualitative Example

alt text