fastfusion

Volumetric 3D Mapping in Real-Time on a CPU

This code implements the approach for real-time 3D mapping on a CPU as described in the following research paper:

http://vision.in.tum.de/_media/spezial/bib/steinbruecker_etal_icra2014.pdf

Volumetric 3D Mapping in Real-Time on a CPU (F. Steinbruecker, J. Sturm, D. Cremers), In Int. Conf. on Robotics and Automation, 2014.

Demo video: http://youtu.be/7s9JePSln-M

@string{icra="Int. Conf. on Robotics and Automation"}
@inproceedings{Steinbruecker-etal-icra14,
  author = {F. Steinbruecker and J. Sturm and D. Cremers},
  title = {Volumetric 3D Mapping in Real-Time on a CPU},
  booktitle = icra,
  year = {2014},
  address = {Hongkong, China},
  titleurl = {steinbruecker_etal_icra2014.pdf},
  topic = {3D Reconstruction},
  keywords =  {RGB-D,Fusion,3d-reconstruction}
}

Installation (Ubuntu 20 LTS)

sudo apt install build-essential cmake libopencv-dev libboost-all-dev libeigen3-dev libglew-dev freeglut3-dev

QT4

As Ubuntu 20 dropped qt4 dependencies in its repos, use rock-core repos :

sudo add-apt-repository ppa:rock-core/qt4
sudo apt update
sudo apt install libqt4-dev

(If Ubuntu 20.10 you might found qt4 via sudo add-apt-repository ppa:gezakovacs/ppa)

QGLViewer

Download and install packages :

Fastfusion

git clone https://github.com/remmel/fastfusion.git
cd fastfusion
cmake .
make

Open with codeblocks: cmake . -G"CodeBlocks - Unix Makefiles"

Preparation of the data

The software takes a text file as input which contains per file

the camera pose
the depth image filename
the color image filename

You can either generate such a file yourself (e.g., by running Christan Kerl's DVO SLAM:

http://vision.in.tum.de/data/software/dvo

available as open source on our homepage) or you can download sequences from the TUM RGB-D benchmark:

http://vision.in.tum.de/data/datasets/rgbd-dataset/

For simplicity, we take a pre-recorded sequence from the TUM RGB-D benchmark.

$ mkdir ~/data
$ cd ~/data
$ wget http://vision.in.tum.de/rgbd/dataset/freiburg3/rgbd_dataset_freiburg3_long_office_household.tgz
$ tar xvzf rgbd_dataset_freiburg3_long_office_household.tgz

Now we need to generate the text file. For this, we use the associate.py tool from the RGB-D benchmark website. We need to run it twice, as we join the camera poses, the depth image list and the color image list into a single file:

$ sudo apt install python3 python3-pip
$ pip3 install numpy
$ cd ~/data/rgbd_dataset_freiburg3_long_office_household/
$ python3 ~/fastfusion/associate.py groundtruth.txt depth.txt > tmp.txt
$ python3 ~/fastfusion/associate.py tmp.txt rgb.txt > associate.txt

The resulting text file should look as follows:

$ head associate.txt

1341847980.790000 -0.6832 2.6909 1.7373 0.0003 0.8617 -0.5072 -0.0145 1341847980.786879 depth/1341847980.786879.png 1341847980.786856 rgb/1341847980.786856.png
1341847980.820100 -0.6821 2.6914 1.7371 0.0003 0.8609 -0.5085 -0.0151 1341847980.822989 depth/1341847980.822989.png 1341847980.822978 rgb/1341847980.822978.png
1341847980.850000 -0.6811 2.6918 1.7371 0.0001 0.8610 -0.5084 -0.0159 1341847980.854690 depth/1341847980.854690.png 1341847980.854676 rgb/1341847980.854676.png
[..]

Running the code

$ ./bin/onlinefusion ~/data/rgbd_dataset_freiburg3_long_office_household/associate.txt --thread-fusion

After some debugging output on the console, a window with a 3D viewer should open. To start the reconstruction process, press "S".

If you run the program for the first time, press and hold the CTRL key and turn your scroll wheel. This is only needed once to "free" the camera viewpoint. After this, you can pan (right click) and rotate (left click) the view as you wish using your mouse.

Further options

   ./bin/onlinefusion  [--intrinsics <string>] [--imagescale <float>]
                       [--threshold <float>] [--scale <float>]
                       [--max-camera-distance <float>]
                       [--consistency-checks <int>] [-k <int>] [-e <int>]
                       [-s <int>] [--incremental-meshing] [-c] [-b] [-v]
                       [--thread-image] [--thread-fusion]
                       [--thread-meshing] [-l <string>] [--] [--version]
                       [-h] <string> ...


Where: 

   --intrinsics <string>
     File with Camera Matrix

   --imagescale <float>
     Image Depth Scale

   --threshold <float>
     Threshold

   --scale <float>
     Size of the Voxel

   --max-camera-distance <float>
     Maximum Camera Distance to Surface

   --consistency-checks <int>
     Number of Depth Consistency Checks

   -k <int>,  --imagestep <int>
     Use every kth step

   -e <int>,  --endimage <int>
     Number of the End Image

   -s <int>,  --startimage <int>
     Number of the Start Image

   --incremental-meshing
     Perform incremental Meshing

   -c,  --loopclosures
     Read Multiple Trajectories and perform Loop Closures

   -b,  --buffer
     Buffer all Images

   -v,  --viewer
     Show a Viewer after Fusion

   --thread-image
     Thread reading the Images from Hard Disk

   --thread-fusion
     Thread the Fusion inside the Viewer

   --thread-meshing
     Thread the Meshing inside the Fusion

   -l <string>,  --loadmesh <string>
     Loads this mesh

   --,  --ignore_rest
     Ignores the rest of the labeled arguments following this flag.

   --version
     Displays version information and exits.

   -h,  --help
     Displays usage information and exits.

   <string>  (accepted multiple times)
     The File Names

UI Options:

S : Start/Stop/Resume fusion
N : Add stop?
G : Fusion next frame
Left : Previous frame
Right : Next frame
W : Write Mesh (.ply)
F : Save Snapshot
O : Resize window
L : Light on/off
C : Color on/off
K : Show depth
P : Mode Faces/WireFrame/Points
D : Save Debug images
Depth image and rgb image must have same resolution and same intrinsics
Enlarging depth png to get a better resolution rgb jpg doesn't change anything

I forked that project in order to use FastFusion reconstruction algo with the rgbd images provided by AREngine on my Honor View 20 using my RGBD Recorder.

remmel / fastfusion