Graphics Reference
In-Depth Information
backprojected
SIFT grid
image plane
scene mesh
Figure 8.33.
Back-projected SIFT features for 3D data, as proposed by Smith et al. [
460
].
in the feature detection and descriptor construction. For example, features that seem
appealing based on the image evidence alone can be ruled out if they straddle a depth
discontinuity in 3D.
Smith et al. later observed that the full invariance of SIFT detection and descrip-
tion isn't necessary for 3D data, since distance measurements produced by LiDAR
already have a physically meaningful scale. That is, even if we view the same object
from different perspectives in two scans, there is no ambiguity about the object's
scale (unlike two images at different perspectives). This insight led to the develop-
ment of
physical scale keypoints
[
459
], which are computed at a predefined set of
physical scales in3D. Unlikeback-projectedSIFT features, the keypoint detectionand
description takes place directly on the 3D mesh, aided by the backprojected texture
from the co-registered images. An analogue of the LoG detection operator is applied
to the texturedmesh, downweighting points whose normals disagreewith the normal
of the point under consideration. A SIFT-like descriptor is computed on the tangent
plane to the detected feature, and descriptors are considered for matching only at
the same physical scale. The overall process eliminates many false matches and has
the additional benefit of allowing the correct detection andmatching of features near
physical discontinuities.
8.4.2
Pairwise Registration
Since LiDARpoints aremeasured in a highly accurate physical coordinate system, any
two scans of the same environment are related by a 3D rigidmotion, that is, a rotation
matrix
R
and a translation vector
t
, each described by three parameters. In some
sense, this is an easier problem than image registration. Only six parameters suffice
to describe the relationship between any two 3D scans, no matter how complex,
while images are only rarely related by simple parametric transformations. On the
other hand, untextured 3D scans are typically much less “feature-rich” than images,
which makes the problemmore challenging.
