Flexible Bump Map Capture From Video
James A. Paterson and Andrew W. Fitzgibbon
University of Oxford
Bump Mapping The Algorithm Calibration
…is a generic term grouping techniques which use 2D texture to •Requirement: Avoid physical measurement
apply 3D detail to a surface. This includes Displacement mapping
•Use set of calibration images of light reflected in mirror, intersect
and our primary interest Normal mapping. We present an improved
ray on which light must lie in at least two images to find relative
method for normal map acquisition from real world objects.
Input Images – A video sequence of the object being moved in
front of a calibrated camera-light rig, or equivalently a portable
camera-light rig moved around whilst viewing a static object
Calibration geometry Images used in calibration
Fronto-parallel Images – By tracking markers on the object we
can compute relative camera-light position, and re-render the •Measurement error makes closed from solution inaccurate. Use a
object using an inverse perspective projection. This is done via 2D non-linear optimisation scheme to optimise camera positions, focal
Standard method Equivalent new method(s)
homography for planar markers. length and light position, minimizing reprojection error:
•Standard normal map capture rig is accurate, but bulky.
•Requires static camera and multiple static lights.
•Proposed new version is portable, only requires one light
•Relative camera-light position is calculated, and must be constant
Sparse Normal Map Images – We can now compute surface
normals as outlined previously. Here we see a sparse set of normals
superimposed on input video frames Current issues and possible extensions
Calculating a surface normal map
•The Lambertian lighting approximation is not perfect, but by
To find the normal at a particular point on the surface, we observe using more than 3 images we can avoid irregular intensity values
the intensity of that point under different lighting conditions. We caused by specular highlights and self shadowing. Determining
model lighting effects via the standard Lambertian equations. exactly which pixels from which images to include in the
Normal map capture requires at least 3 different intensity values computation is non-trivial, but we would hope to maximise on the
for a point on the surface, captured with different known light range of light directions used in calculating each normal.
directions. Then we can find the normal at that point as follows:
•Currently only planar surfaces can be captured, however by
Dense Normal Map Images – A higher density grid superimposed improving the geometrical model of the subject we hope to extend
on a longer sequence of input images as a bump mapped the process to highly curved surfaces.
quadrilateral. Non-Lambertian lighting conditions (specular
•The positions of the markers on the subject must be accurately
highlights, self shadowing) lead to occasional spurious results.
known, however in e.g. archive footage this information will not be
Limitation: Planar surfaces only at the moment! available. Given sufficient images we can include marker positions
in the non-linear optimisation, removing this restriction