CAVIAR Model Selection

We developed a Bayesian Model Network that fuses information about both target interestingness (newness, priority area) and model interestingness (fighting is more interesting than walking). Additional information is the conclusion for the hypothesis's context in the previous frame. The module ranks the different context hypotheses that could be applied to the different tracked targets. This gives a priority list that the situation/context recogniser can use to prioritise its computational resources. In the end, we did not integrate this module, in part because the situation/context recogniser actually ran quite quickly compared to some of the earlier modules in the CAVIAR system (e.g. dense feature extraction).

Experiments showed that, when using the previous frame's results, the average number of false hypotheses considered before each target's true hypothesis was encountered was 0.09 (out of 7). When not using the previous frame's results, then, on average, 1.2 false hypotheses (out of 7) were considered before the true hypothesis. Random hypothesis selection would normally cause 3.5 hypotheses to be considered on average, so both approaches give much better results than random selection.

A paper that describes the process is:

  1. d23.pdf

Back to CAVIAR home page.