GENERAL INFORMATION
-------------------

This package contains scan data as used in our paper [1] for fully automatic pairwise registration
of wide-baseline scans.

A scan consists of both, depth and image data. In our case depth was measured via a laser scanner
(Leica C10) and images were captured with the scanner's built in camera or an external camera
mounted on top of the scanner. In either case images as well as the 2.5D scene structure are
represented as 6 faces of a cube map with depth and RGB camera sharing the same single center of
projection. This RGB-D cube-maps are considered as input for our registration algorithm in [1].

Our project web page is http://cvg.ethz.ch/research/saldir-rgbd-registration/.

For questions please contact me via zeislb__at__inf.ethz.ch.


CONTENT OF THE PACKAGE
----------------------

Each SCAN contains following data:

1) scan.pcd is the original point cloud as obtained by the scanner. *.pcd is a file format as used
by the pointcloud library (PCL), http://www.pointclouds.org/, and can e.g. viewed via pcl_viewer
scan.pcd after installation of the library.

2) cube_depth_[00-05].mat contain the 6 cube face depth-maps in floating point precision. They are
generated from the original point cloud. Note that *.mat is NOT a Matlab data type, but a simple
binary file storing data of a 2D grid in row-major order. It has an 8 byte header (2 integers)
storing height and width, following by height*width float entries encoding the depth values in
meters. The provided Matlab script readMatrix.m reads such files into Matlab.
If you view the depth maps sometime you will observe thin vertical structures in front of facades,
buildings, etc. They originate from people which were walking through the scene while we were
scanning. Since the scanning process is slow, only a view vertical sweeps were actually hitting a
person walking by. In our algorithm we (automatically) treat these spikes as noise during
orthographic rendering.

3) cube_depth_[00-05].png contain images of the 6 cube face depth-maps. This is just for
visualization. Use the *.mat file to obtain the accurate depth measurements.

4) cube_image_[00-05].png are the 6 cube face images (24 bit). In general they do not have the same
resolution as the depth maps, but are larger. Still a pixel-to-pixel RGB-to-depth mapping is simply
achieved by appropriate scaling.

5) cube.wrl is a simple 3D model file mapping the 6 cube face images onto a cube for visualization
of the scene. E.g. the free Instant Player from Instant Reality (http://www.instantreality.org/) can
be used for viewing.

In addition each SCENE contains:

1) cube_scene_camera_models.txt lists the intrinsic and extrinsic parameters for each cube face
camera. Basically viewpoint ID 0 has pose [I,0] and the remaining poses are obtained by 90 degree
rotations of the camera. For all cameras (depth and image) the center of projection is identical in
the center of the cube.

2) registration_result.pcd is the aligned point cloud we obtain with our approach [1]. The point
cloud contains all scans (scan.pcd) fused into one model and subsampled for smaller file size. In
addition points are colored with the image pixel color they project into. Again *.pcd is the PCL
file format.


RELATED PUBLICATIONS
--------------------

[1] Automatic Registration of RGB-D Scans via Salient Directions
    Bernhard Zeisl, Kevin Koeser, Marc Pollefeys
    Int. Conf. on Computer Vision (ICCV), 2013, Sydney, Australia 


---
http://cvg.ethz.ch/research/saldir-rgbd-registration/
v1.0, Bernhard Zeisl, Computer Vision and Geometry Group, ETH Zurich, 12.12.2013