/rpg_emvs

Code for the paper "EMVS: Event-based Multi-View Stereo" (IJCV, 2018)

Primary LanguageC++OtherNOASSERTION

EMVS: Event-based Multi-View Stereo

EMVS: Event-based Multi-View Stereo

This is the code for the 2018 IJCV paper EMVS: Event-Based Multi-View Stereo - 3D Reconstruction with an Event Camera in Real-Time by Henri Rebecq, Guillermo Gallego, Elias Mueggler, and Davide Scaramuzza.

Citation

A pdf of the paper is available here. If you use any of this code, please cite this publication as follows:

@Article{Rebecq18ijcv,
  author        = {Henri Rebecq and Guillermo Gallego and Elias Mueggler and
                  Davide Scaramuzza},
  title         = {{EMVS}: Event-based Multi-View Stereo---{3D} Reconstruction
                  with an Event Camera in Real-Time},
  journal       = "Int. J. Comput. Vis.",
  year          = 2018,
  volume        = 126,
  issue         = 12,
  pages         = {1394--1414},
  month         = dec,
  doi           = {10.1007/s11263-017-1050-6}
}

Patent & License

  • The proposed EMVS method is patented, as you may find in this link.

    H. Rebecq, G. Gallego, D. Scaramuzza
    Simultaneous Localization and Mapping with an Event Camera
    Pub. No.: WO/2018/037079.  International Application No.: PCT/EP2017/071331
    
  • The license is available here.

Overview

From a high-level, input-output point of view, EMVS receives a set of events and camera poses and produces a semi-dense 3D reconstruction of the scene, as shown in the above video. See the example below.

Installation

This software depends on ROS. Installation instructions can be found here. We have tested this software on Ubuntu 16.04 and ROS Kinetic.

Install catkin tools, vcstool:

sudo apt-get install python-catkin-tools python-vcstool

Create a new catkin workspace if needed:

mkdir -p ~/emvs_ws/src && cd ~/emvs_ws/
catkin config --init --mkdirs --extend /opt/ros/kinetic --merge-devel --cmake-args -DCMAKE_BUILD_TYPE=Release

Clone this repository:

cd src/
git clone git@github.com:uzh-rpg/rpg_emvs.git

Clone dependencies:

vcs-import < rpg_emvs/dependencies.yaml

Install pcl-ros:

sudo apt-get install ros-kinetic-pcl-ros

Build the package(s):

catkin build mapper_emvs
source ~/emvs_ws/devel/setup.bash

Running example

Download slider_depth.bag data file, from the Event Camera Dataset, which was recorded using the DVS ROS driver.

Run the example:

roscd mapper_emvs
rosrun mapper_emvs run_emvs --bag_filename=/path/to/slider_depth.bag --flagfile=cfg/slider_depth.conf

Configuration parameters: The options that can be passed to the program using the configuration file (e.g., slider_depth.conf) and their default values are defined at the top of the main.cpp file. These are: the parameters defining the input data, the parameters of the shape and size of the Disparity Space Image (DSI), and the parameters to extract a depth map and its point cloud from the DSI.

Visualization

Images

Upon running the example above, some images will be saved in the folder where the code was executed, for visualization. For example, the output images for the slider_depth example should look as follows:

Confidence map Depth map

The depth map is colored according to depth with respect to the reference view.

Disparity Space Image (DSI)

We also provide Python scripts to inspect the DSI (3D grid).

Volume Rendering

Install visvis first:

pip install visvis

To visualize the DSI stored in the dsi.npy file, run:

roscd mapper_emvs
python scripts/visualize_dsi_volume.py -i /path/to/dsi.npy

You should get the following output, which you can manipulate interactively:

Showing Slices of the DSI

To visualize the DSI with moving slices (i.e., cross sections), run:

python scripts/visualize_dsi_slices.py -i /path/to/dsi.npy

which should produce the following output:

Point Cloud

To visualize the 3D point cloud extracted from the DSI, install pypcd first as follows:

pip install pypcd

and then run:

python scripts/visualize_pointcloud.py -i /path/to/pointcloud.pcd

A 3D matplotlib interactive window like the one below should appear, allowing you to inspect the point cloud (color-coded according to depth with respect to the reference view):

Additional Examples

We provide additional examples with sequences from the Event Camera Dataset.

Office Scene

Download dynamic_6dof and run:

rosrun mapper_emvs run_emvs --bag_filename=/path/to/dynamic_6dof.bag --flagfile=cfg/dynamic_6dof.conf

The images generated should coincide with those in this folder.

Confidence map Depth map

You may also explore the DSI as in the previous example (the same commands should work).

Boxes

Download boxes_6dof and run:

rosrun mapper_emvs run_emvs --bag_filename=/path/to/boxes_6dof.bag --flagfile=cfg/boxes_6dof.conf

The images generated should coincide with those in this folder.

Confidence map Depth map

You may also explore the DSI as in the previous example (the same commands should work).

Shapes

Download shapes_6dof and run:

rosrun mapper_emvs run_emvs --bag_filename=/path/to/shapes_6dof.bag --flagfile=cfg/shapes_6dof.conf

The images generated should be those in this folder.

Confidence map Depth map

As you may notice by inspecting the DSI, the shapes are on a plane (a wall).

Additional Notes

By default, the Z slices of the DSI are uniformly spaced in inverse depth. However, it is possible to change this behavior to use Z slices uniformly spaced in depth (rather than inverse depth). This can be achieved by changing the option USE_INVERSE_DEPTH to OFF in the CMakeLists.txt. This requires recompiling mapper_emvs. We recommend removing the emvs_mapper build folder before recompiling.

Additional Resources on Event Cameras