Learning attention models for resource-constrained, self-adaptive visual sensing applications.

Resource constraints are one of the main design challenges for wireless sensor network applications and visual sensing networks that employ cameras in particular. The objective in this paper is to enable the sensors to be context-aware by utilizing application-level information, to prioritize parts of an image, and only transmit those parts that contribute most to the utility of the application. We, therefore, study online-learning of visual attention models for the use case of person detection and counting. We analyze how the resulting models can prioritize relevant elements of a partial image, so that object detection remains accurate compared to a random selection strategy when resources for transmission get scarce. Results show that such attention models can be learned also under constraints and converge towards the true models. For the application performance, we observed an average reduction of errors (the number of undetected persons) of 55% compared to policies without a corresponding attention model.