ROST is a realtime online spatiotemporal topic modeling framework for data such as streaming video and audio observed by a robot, where topics represent the latent causes that produce these observations. When new observations are made, we not only compute its topic labels, but also use it to update the global topic model and the labels of older observations, resulting in a consistent semantic description of the scene without the use of any prior knowledge. The proposed approximations have constant update time as new data is observed, which allows for the technique to work in real-time; a critical requirement for its use in the robotics.
http://cim.mcgill.ca/~yogesh/rost/
