Information Integration and Belief Inconsistency

The solutions to many vision problems involve integrating measurements from multiple sources. Most existing methods assume these measurements to be consistent. In reality, unfortunately, this assumption may not hold and naively fusing inconsistent measurements is very likely to fail. In our recent study, we proposed a novel approach to handling the inconsistency. When representing the multiple sources as a random field, we find that the belief inconsistency is closely related to the fixed-points of the field and prove a new theorem that gives two algebraic criteria for evaluating the inconsistency. Based on this theoretical analysis, we provide a new information integration method that leads to very encouraging results in visual tracking tasks.

Demo

Belief inconsistency in visual tracking (13M including many examples)

The estimation of the target (i.e., the head) depends on the integration of the estimates of many facial components (i.e., various information sources). In this video, some of the components are occluded when the target moves, and their estimates are wrong. As a consequence, their corresponding information sources may provide inconsistent or even conflicting information to the integration process. An effective method is needed to detect and resolve this inconsistency (for details please see our papers).

Publication:

Gang Hua, Ying Wu and Zhimin Fan, "Measurement Integration Under Inconsistency for Robust Tracking", in Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR 06), New York City, NY, June 17-22, 2006. [PDF]

Return to Tracking Research