PennOffice: A Room Reconstruction from Quadrinocular Stereo Vision

Radim Sara and Ruzena Bajcsy

GRASP Laboratory, Department of Computer Science, University of Pennsylvania
sara@cmp.felk.cvut.cz, http://cmp.felk.cvut.cz/~sara/home.html

Four-camera stereo setup was used to reconstruct about 150 degrees wide view of a 3 x 4.3 m office from the viewpoint of a sitting person. An additional, fifth, narrower-field-of-view camera was used to acquire 24-bit color images for color mapping.

wide-angle view of the scene
A wide-angle view of the scene. Two cameras of the stereohead are visible in the foreground.

3-D Euclidean reconstruction

The reconstruction we have been able to get from our image data is a set of colored points in Euclidean 3-D space. The sets are automatically registered with about 1cm accuracy. These examples show 80% of the reconstructed data (segments A and C) comprising partial reconstructions from 40 views.

view 5 view 1 view 6
view 3 view 10 view 12
view 11

Demos

What do we want to do with it?

We have collected a large dataset of image data from carefully calibrated cameras. The dataset could become a testbed for testing stereo or structure-from-motion algorithms and for scenes modeling.

The data contain two sets of images of exactly the same scene, one with natural surface texture and one with enforced (projected) random texture. Reconstructions from natural images can be compared against reconstructions from ideal images.

The reconstructed scene is a collection of colored isolated points in Euclidean space, each point is a six-tuple [x, y, z, r, g, b], see Reconstruction Datafiles.

In addition to the reconstructed scene, we can provide all the camera calibration parameters and all the motion parameters but we can also provide all the images, including those used for camera calibration and motion estimation. We can provide calibration, motion estimation, and reconstruction procedures we used (in Matlab and C code).

All of this will be available soon. Stay tuned.

Pages Contents

  1. Experiment Description
  2. Input Images and Camera Calibration Data
  3. Reconstructed 3-D Point Models for Matlab®
  4. Reconstructed 3-D Point Models for OpenInventor
Radim Sara
Last modified: Thu May 7 09:32:16 MET DST 1998