Computing Epipolar Geometry From Dynamic Silhouettes

1
Camera Network Calibration from Dynamic
Silhouettes Sudipta N. Sinha
Marc Pollefeys Leonard
McMillan Department of Computer Science,
University of North Carolina at Chapel Hill
Results Epipolar Geometry computed using
proposed approach

Corresponding
Frames in 2 views
and manually
clicked points for
UNC dataset (left)
and Keck Lab
dataset (see below).
Goal To Calibrate a Network of Cameras from
video streams.
Results Full Calibration from 4-view MIT
dataset.
View 1
P1 P2 . . Pn
View 2
Metric Calibration Of Camera Network From
Fundamental Matrices
F12 . . Fjk
Compute Pairwise Epipolar Geometry using
Silhouettes
4
Silhouette Extraction
3
Input Video Sequences
1
2

• Advantages
• No Calibration Data or special objects(LED,
  pattern) required.
• Handles arbitrary camera configuration and
  wide-baselines.
• Works when corresponding scene features are few
  or unavailable,
• without photometric calibration and when image
  quality is poor.

View 2
View 1
View 4
View 3
Computing Epipolar Geometry From Dynamic
Silhouettes
Keyframes were selected from 4 min. long, 30 fps
video sequences
Calibration, Visual Hull and Reprojection
Test (Below) The calibration was used to
construct the visual hull. (Bottom) The Visual
Hull Reprojection Error in the 4 views.
RMS Error 1.38 pixels
RMS Error 1.31 pixels
Metric Camera Calibration from Fundamental (F)
Matrices.
Frontier Points and constraints provided by
Epipolar Tangents.
N-View Graph G(V,E) V set of N camera
views. E set of F matrices computed by our
method.
Tangent Envelope This is a compact representatio
n for silhouettes, allows efficient use of
long video streams.
RANSAC-based Algorithm If the epipoles are
known, then 3 pair of corresponding lines through
the epipoles are required to compute the
epipolar homography. RANSAC is used to
randomly explore the 4D space of epipole
hypothesis as well as robustly deal with
incorrect silhouettes.
Computing Projective Cameras Building
Block Triplet of Views. Given, F12, F23, F13 A
consistent set of Projective Cameras P1, P2,
P3 are computed. Method of Induction used
to add New views to existing network.
Need Sub-frame synchronization.
Hypothesis Step

• Work in progress (papers in upcoming
  conferences)
• S. Sinha, M. Pollefeys. Synchronization and
  Calibration of Camera
• Networks from Silhouettes, ICPR 2004.
• S. Sinha, M. Pollefeys. Visual-Hull
  Reconstruction from Uncalibrated
• and Unsynchronized Video Streams, 2nd
  International Symposium
• on 3D Data Processing, Visualization
  Transmission, 2004.
• Currently exploring Active PTZ Camera Network
  Calibration.

Applications 3D Event Reconstruction in
Surveillance Systems.
Verification Step