Real-time motion from structure
Switch branches/tags
Nothing to show
Clone or download
Permalink
Failed to load latest commit information.
gif first commit Oct 7, 2018
rpg_svo first commit Oct 7, 2018
rpg_vikit first commit Oct 7, 2018
.gitignore first commit Oct 7, 2018
README.md updated the links to paper and video Oct 7, 2018

README.md

CNN-SVO

(Top) camera frames (Bottom) depth maps of the keyframes from Monodepth

Demo: https://www.youtube.com/watch?v=4oTzwoby3jw

Paper: CNN-SVO: Improving the Mapping in Semi-Direct Visual Odometry Using Single-Image Depth Prediction

This is an extension of the SVO with major improvements for the forward facing camera with the help of monocular depth estimation from CNN.

We tested this code with ROS Kinetic on Ubuntu 16.04.

There are two ways to make it work

  1. Online mode
  2. Offline mode

Dataset

We use KITTI Odometry sequences and Oxford Robotcar dataset to evaluate the robustness of our method.

We provide camera parameters for both datasets. Running a sequence requires setting the directory that keeps the images (also directory to their corresponding depth maps if you are using Offline mode).

Unfortunately, getting Oxford Robotcar dataset to work is not straightforward. Here is the rough idea of what do you need to do to preprocess Oxford Robotcar dataset images:

  1. Pick a sequence to download
  2. Use the SDK to get the RGB images
  3. Crop the images to match the aspect ratio (1248x376) of images in KITTI dataset
  4. (For running Offline mode) Save the depth maps of those images

Online mode

  1. Clone the monodepth-cpp repo and follow the instructions
  2. Make sure the library is successfully built by running the inference_monodepth executable
  3. Set the image directory in rpg_svo/svo_ros/param/vo_kitti.yaml
  4. Make sure the images are colour (not greyscale) images

NOTE: If you are having difficulty compiling the library, you can still visualize the results using the Offline mode

Offline mode

  1. Clone the Monodepth repo and follow the Testing instructions
  2. Modify the monodepth_main.py so that it saves the disparity maps as numpy (.npy) files in one folder (per sequence)
  3. Set the image and depth directories in rpg_svo/svo_ros/param/vo_kitti.yaml

NOTE: For saving disparity map, the naming convention is depth_x.npy, where x is 0-based with no leading zero.

Instructions

  1. Clone this repo
cd ~/catkin_ws/src
git clone https://github.com/yan99033/CNN-VO
  1. (OPTIONAL) Clone monodepth-cpp repo
cd ~/catkin_ws/src/CNN-VO
git clone https://github.com/yan99033/monodepth-cpp
  1. Enable/disable Online mode by toggling TRUE/FALSE in svo_ros/CMakeLists.txt and svo/CMakeLists.txt. Also make sure that Monodepth library and header file are linked properly, if Online mode is used. Note that Online mode is disabled by default

  2. Compile the project

cd ~/catkin_ws
catkin_make
  1. Launch the project
roscore
rosrun rviz rviz -d ~/catkin_ws/src/CNN-VO/rpg_svo/svo_ros/rviz_kitti.rviz
roslaunch svo_ros kittiOffline00-02.launch
  1. Try another KITTI sequence
  • Set the folder image and depth directories in rpg_svo/svo_ros/param/vo_kitti.yaml
  • Use the corresponding roslaunch file

Disclaimer

The authors take no credit from SVO and Monodepth, therefore the licenses should remain intact. Please cite their work if you find them helpful.