Fast Robust Large-scale Mapping from Video and Internet Photo ...

Building image pyramids and tracking features are both highly parallel
operations and so can be programmed efficiently on the graphics processor
(GPU). Building an image pyramid is simply a series of convolution operations
with Gaussian blur kernels followed by down sampling. Differential
tracking can be implemented in parallel by assigning each feature to track to
a separate processor. With typically on the order of 1024 features to track
from frame to frame in a 1024x768 image we can achieve a great deal of
parallelism in this way.
One major limitation of standard KLT tracking is that it uses the absolute
difference between the windows of pixels in two images to calculate
the feature track update. When a large change in intensity occurs between
frames this can cause the tracking to fail. This happens frequently in videos
shot outdoors when the camera moves from light into shadow for example.
Kim et al. in [62] developed a gain adaptive KLT tracker that measures
the change in mean intensity of all the tracked features and compensates for
this change in the KLT update equations. Estimating the gain can also be
performed on the GPU at minimal additional cost in comparison to standard
GPU KLT.
4.1.2. Robust Pose Estimation
The estimated local correspondences typically contain a significant portion
of erroneous correspondences. To determine the correct camera position
we apply a Random Sample Consensus (RANSAC) algorithm [63] to estimate
the relative camera motion through the essential matrix[20]. While
being highly robust the RANSAC algorithm can also be computationally
very expensive, with a runtime that is exponential in outlier ratio and model
14