performance of VSLAM

edit

When the camera is moving quickly (particularly when it's rotating), every stage of VSLAM gets harder. It's harder to extract features, because of camera blur. It's harder to match features, because it's hard to initialize well, and because the viewpoint has changed, fouling up which features are visible. The geometric constraints that can help start to fall apart in the monocular case with rotations.

In general, VSLAM (and monocular VSLAM in particular) are very tricky; one of the reasons the VSLAM package isn't getting attention is because it's very, very hard to make VSLAM systems work reliably for everyone; a package that works great on a PR2, in a particular environment, may need to be re-tuned arbitrarily for your robot in your building; this just isn't something where general-purpose code that Just Works (or even Sorta Just Kinda Works) can be shipped.