View Selection

Operational steps involved in best-camera selection


	This page provides the results for various camera selection algorithms. For various related publications visit the Publications webpage.
Camera selection
	Best camera selection ( $Υ$ _max) A naive method of selecting the best-camera given the QoV would be to select the best-camera at each time. This naive method of selecting the best camera at each time instant does not typically yield visually pleasant videos. Noisy measurement cause erroneous camera selection decisions and result in poor synthesized video with frequent inter-camera switching. A video showing this switching is presented below.

	Camera scheduling using state modeling ( $Υ$ _dbn) The view-selection process can be aided via the use of scene-centric state modeling. This modeling is based on a Dynamic Bayesian Networks (DBN) to integrate camera network information as a prior for selection. DBN allows us to achieve an optimal number of camera switches by enforcing temporal smoothing. It employs previous view information to select the current view thus increasing resilience to frequent switching. A video comparing the output of this method with the best-camera selection using $Υ$ _max is shown below.

	Camera scheduling using utility maximization ( $Υ$ _util) We can also model the view-selection problem as a decision process during which information is only partially visible. In particular we use a Partially Observable Markov Decision Process (POMDP), where the process measured by the cameras (e.g., object size, location or scene activity) is a Markov process and the sensor scheduling is based on recursively estimating and updating the belief state, the sensor-scheduling actions, and the posterior distribution of the process given the history of the sensor measurements. We represent the process dynamics and measurements as linear Gaussian state-space models and track the belief using the Bayes Rule. The reward for camera selection is modeled as a function on a content-quality score and the related camera switching. This reward modeling allows the proposed approach to control the number of camera switches, thus enabling the generation of pleasant videos.

	Comparison Both the proposed $Υ$ _dbn and $Υ$ _util estimate the next best-camera based on information from the past. However $Υ$ _util models the repercussions of this selection on the future decissions as well. A video showing the comparison of the proposed $Υ$ _dbn, $Υ$ _util with the manually generated ground truth $Υ$ _gt is shown below.

Multi-camera scheduling

Camera selection