Pareto Optimization of Parameter Selection Speeds Up and Improves Quality of Motion Computation: Applying Evolutionary Multi-objective Optimization to Randomized-Subspace Robust PCA
David Grob, Mehmet Vurkaç, Agnieszka C. Miguel, Mirka M. Mandich, Rana Bayrakçısmith · 2018
Camera trapping is a noninvasive tool to collect data about wildlife. For example, conservation biologists use camera traps to study the elusive snow leopard. Widespread adoption of camera traps has resulted in a large volume of image data produced every day. There is a pressing need to develop automated analysis methods that can free the researchers from long hours of manual labor. In this work, we show a new method to extract motion from sets of consecutive images. We introduce significant improvements in speed and accuracy by applying multi-objective optimization to the determination of parameters used in motion computation via randomized-subspace RPCA. Our results show improvement on holdout data even when the content of image sets differs considerably. Between uncorrelated image sets, we improve execution speed by up to 24% on some of the holdout data, and the quality by as much as 53% as compared to an ad hoc choice of parameters.