Closed ahangchen closed 4 years ago
You are right about the assumption. When the camera is rotating, optical expansion computed from optical flow fields is not equivalent to motion-in-depth.
However, the camera rotation is relatively easy to remove by rotational homograph (assuming you already have R from visual/inertial odometry methods)
Got it, thank's for your reply.
In 3.1, there is an assumption that orientation is unchanged, but when camera's rotation is not identity, the orientation will change (for example, a car with camera turns around at road corner). Is there anything I misunderstand about orientation?