Just Go with the Flow: Self-Supervised Scene Flow Estimation

Code release for the paper Just Go with the Flow: Self-Supervised Scene Flow Estimation, CVPR 2020 (Oral).

Authors: Himangi Mittal, Brian Okorn, David Held

[arxiv] [Project Page]


If you find our work useful in your research, please cite:

author = {Mittal, Himangi and Okorn, Brian and Held, David},
title = {Just Go With the Flow: Self-Supervised Scene Flow Estimation},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}


In this work, we propose a method of scene flow estimation using two self-supervised losses, based on nearest neighbors and cycle consistency. These self-supervised losses allow us to train our method on large unlabeled autonomous driving datasets; the resulting method matches current state-of-the-art supervised performance using no real world annotations and exceeds stateof-the-art performance when combining our self-supervised approach with supervised learning on a smaller labeled dataset.

For more details, please refer to our paper or project page.



CUDA 9.0  
Tensorflow-gpu 1.9
Python 3.5
g++ 5.4.0


(a). Clone the repository.

git clone https://github.com/HimangiM/Self-Supervised-Scene-Flow-Estimation.git

(b). Install dependencies

Create a virtualenv
python3 -m venv sceneflowvenv
source sceneflowvenv/bin/activate
cd Self-Supervised-Scene-Flow-Estimation
pip install -r requirements.txt

Check for CUDA-9.0

(c). Compile the operations The TF operators are included under src/tf_ops. Check the CUDA compatability and edit the architecture accordingly in makefiles of each folder (tf_ops/sampling, tf_ops/grouping, tf_ops/3d_interpolation) The authors had used sm_61 as the architecture for CUDA-9.0. Finally, move into each directory and run make. Also, check for the path for CUDA-9.0 and edit the path in the makefiles of each folder. If this method throws error, then run bash make_tf_ops.sh sm_61.


Download the kitti dataset from the Google Drive link. Each file is in the .npz format and has three keys: pos1, pos2 and gt, representing the first frame of point cloud, second frame of point cloud and the ground truth scene flow vectors for the points in the first frame. Create a folder with name data_preprocessing and download the kitti dataset in it. The dataset directory should look as follows:

|  |--kitti_self_supervised_flow
|  |  |--train
|  |  |--test

The data preprocessing file to run the code on KITTI is present in the src folder: kitti_dataset_self_supervised_cycle.py. To create a dataloader for own dataset, refer to the script:


Training and Evaluation

To train on own dataset, refer to the scripts:

bash src/commands/command_train_cycle_nuscenes.sh

To evaluate on the KITTI dataset, execute the shell script:

bash src/commands/command_evaluate_kitti.sh

Link to the pretrained model.


You can use Open3d to visualize the results. A sample script is given in visualization.py