Discuss

[shadow1runner]

• Influence of too low random sampling (mostly experienced with 'random' effects in the unit test suite, results differ from run to run)

[shadow1runner]

• Various optical flow algirhtms, see here
  • Performance -> GPU, very slow

[shadow1runner]

• also see issue #55: Elaborate on the choice of the measurement noise covariance matrix R:

R = [(1/IR) 0 
         0 (1/IR)]

with IR being the inlier ratio of the FOE

[shadow1runner]

• Unit Test Infrastructure

[shadow1runner]

• Differences in configuration between different camera types to achieve results
  • subsampling (depends on resolution), desired resolution: 180x120 (current configuration, as of 31.8.16 at least)
  • ???

[shadow1runner]

• Different algorithms to solve the set of linear equations using the table here -> use performance and timing measurement and compare colPivHouseholderQr with householderQr.

Discuss:

Unlike \cite{stabinger}, these linear equations are solved for $\bm{x}$ using a QR decomposition, a subtle difference that boosts not only performance but only the numeric quality of the divergence estimation as discussed later in \ref{cha:discussion}.

[shadow1runner]

• Choice of not skipping any frames -> performance is high enough for real-time