Temporal Scene Flow (ICRA 2017)
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.
OpenSceneGraph @ f4013a6


This is my research code for the paper "A Learning Approach for Real-Time Temporal Scene Flow Estimation from LIDAR Data" presented at ICRA 2017. Please contact me (aushani@gmail.com) with any questions or comments. Please cite the paper if this code is useful in your work. (bibtex)

youtube video

Getting Started

If you have any questions, please contact aushani@gmail.com. Please note that the models here (background filter, occupancy constancy) are trained for the KITTI dataset.


Cuda is assumed to be already installed (you need a GPU to run the code). Other dependencies are bazel, Qt4, and OpenSceneGraph. You also need the KITTI data. The bootstrap.sh script takes care of these except for cuda.


To get started, run:

$ ./bootstrap.sh

You will be prompted for:

  • Bazel

    Build system to compile the project

  • Qt4

    from apt-get

  • OpeneSceneGraph

    Used for rendering, locally built

  • KITTI Data

    Prepackaged in the directory structure that the code is expecting. You also have the option of downloading a small snippet of it if bandwidth or diskspace is a concern.

Compile the Project

To compile everything, go to the cc directory and build with bazel:

$ cd cc
$ bazel build //...


After building, you can run with a command like:

$ bazel-bin/app/flow/viewer

This assumes that the data was put in ~/data/tsf_data. Otherwise:

$ bazel-bin/app/flow/viewer --tsf-data-dir path/to/tsf_data