/sac-flow

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

Primary LanguagePython

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

Installation

  • Install Fast Geodis with pip install FastGeodis --no-build-isolation
  • Install PyTorch3d with CUDA support.
  • Run commands in install.sh for installation of the packages above

DATA

  • Setup directory for extracting the data, visuals and experimental results
export BASE_PATH='path_where_to_store_data'
  • Download Data and unpack it to the folder $BASE_PATH/data/sceneflow:
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:

Dataset Argument Number Model
KITTI t 0 Neural Prior
StereoKITTI 1 Neural Prior
KITTI t 2 SCOOP
StereoKITTI 3 SCOOP
Argoverse 4 Neural Prior
Nuscenes 5 Neural Prior
Waymo 6 Neural Prior

Experimental results on LiDAR Datasets

Experimental results on StereoKITTI Dataset

Qualitative Example

alt text