Graphics Reference
In-Depth Information
Fig. 2.3
View recognition process
by view. At this point, crude structure constraints can be applied, using a Hough trans-
form or RANSAC to enforce a loose similarity, affine, homography, or reprojection
transformation. The views are then ranked by the number of matches satisfying the
constraints. The
m
highest ranked views are kept as candidates.
2.6.2 Robust Matching and Pose Estimation
For each of the
m
candidate views chosen by the appearance matching stage, the
query features are matched to the view's features using the local view appearance
model. Each view feature has an associated three-dimensional structure estimate,
allowing the three-point pose algorithm [
6
] to be applied within a standard RANSAC
hypothesize-and-test framework. If enough inliers result from this process, they are
passed to the pose estimation stage, along with the relative pose estimate given by
the three-point algorithm.
The pose estimation stage takes correspondences between query features and
view features, and computes the relative camera pose between the query image and
the view's base coordinate frame (the first image of the view pair). The relative pose
