Details on the underlying theory can be found in
J. Balzer, M. Peters, S. Soatto: Volumetric Reconstruction Applied to Perceptual Studies of Size and Weight, IEEE WACV, 2014.
A preprint is available here. Please cite if you use the software in your own published work.
- acquisition and wide-baseline alignment of RGBD images
- elaborate sensor model
- Poisson mesh reconstruction
- ground plane estimation
- supports all Primesense/ONI devices
- acquisition, import, and export of raw ONI data streams
- QT frontend
- file I/O of
- raw single-view RGB and disparity data (OpenEXR,PNG,PGM)
- estimated poses (TXT or embedded in OpenEXR)
- oriented point clouds (PLY)
- normal maps (R4R)
- reconstructed triangular mesh (PLY)
- 3d viewer
Installing the driver
Unless you have upgraded your Ubuntu to version 13.04 or 13.10, you should be fine with the precompiled binaries which can be download from the Primesense website. Simply unzip the driver and call the install script, while the sensor is not connected to your machine. Once the driver is installed, a good test to verify whether it is working properly is to plug in your device and run the
SimpleViewer located in the subfolder
Otherwise, you will need to download the sources from GitHub. Inside the make file
pthread to the variable
$USED_LIBS in line 29. From inside the root directory call
/Bin/x64-Releaseto it. Run
Clone the repository, e.g., by typing
git clone https://bitbucket.org/jbalzer/yas.git
yas. Change into this directory
cd build qmake .. "OPENNI_DIR=/home/jbalzer/Workspace/OpenNI"
$OPENNI_DIR/Redistto the folder that contains the YAS executable, e.g., from within the build directory type
cp -ra /home/jbalzer/Workspace/OpenNI/Redist/. .
A note for Mac users
The build process on MacOS is slightly more involved, the reason being that OpenEXR relies on POSIX semaphores for thread synchronization. Darwin, the basis of MacOS, is not 100% POSIX compliant in that it does not support semaphores. Any attempt to link against OpenEXR binaries included as part of OpenCV or obtained e.g. through MacPorts will cause a runtime error on EXR export. If you are working on MacOS you will have to build the OpenEXR library from source with semaphore support deactivated.
The basic reconstruction workflow is as follows:
- Start the live stream by clicking Run in the main window or load previous raw measurements stored in EXR format (see below).
Click on Store to add an RGBD image to the sparse set used for reconstruction. Make sure that there is enough overlap between two views. In our experience, roughly eight images taken from all around the object are sufficient for a complete reconstruction. Yet, the common wisdom applies that the smaller the baseline (in terms of translation and rotation), the likely it is that the alignment will be successful. Do not worry about holes. They will be filled in during Poisson reconstruction. The spin box in the lower left lets you iterate through all acquired images, the respective depth and RGB images being displayed in the lower half of the main window.
Hint: Save and re-load raw data to EXR format with File:Save all... respectively File:Open.
Pairwise alignment is triggered (e.g., for parameter tuning) by selecting Align:Current from the Tools menu. This affects the view which is currently on display, say view number i, and the next one, i.e., view number i+1. If the last acquired image is selected, there will be no effect. The correspondence estimation and ICP refinement may take a little bit of time so be patient. A batch alignment is requested by selecting Align:All.
- You can verify the success of the last step by opening the OpenGL window (View:3D View...) and by iterating through the images with the spin box. That is because all point clouds are rendered w.r.t. the world (i.e., the first view's) reference frame. Additionally, upon completion of the alignment, the percentage of inlying matches is displayed in the status bar at the bottom of the main window. If the percentage drops below 10%, the alignment has probably failed. In that case, adjust the SIFT parameters (View:Preferences) or enrich the scene with textured objects, e.g., place the object on a newspaper or similar. If necessary, you can selectively remove views by hitting Clear. This affects the currently selected image.
- Unwanted clutter in fore- and background can be removed by operating the Max/Min sliders at the bottom of the main window. Inspect the cropping result in the bottom depth visualization.
Tools:Poisson creates a triangle mesh from the point cloud. File:Save with the combo box set to PLY now saves this mesh instead of the oriented point cloud currently on display.
Caveat: At the present time, the third-party open-source code for Poisson reconstruction does not perform as well as the one integrated in Meshlab. We will investigate the issue and fix it in future releases. For the time being, it is recommended to export the aligned oriented point clouds to PLY (choose Save all... from the File menu, enter a prefix, select PLY from the combo box in the lower right of the dialogue and hit OK), import them to Meshlab, merge visible layers (right-click on the layers panel, don't forget to check Keep unreferenced vertices), and finally start mesh reconstruction with Filter:Point Set:Surface Reconstruction:Poisson.
A more detailed tutorial on the basic usage of YAS including a screencast is coming soon.