RI-Small: Optimal Automated Design of Cascaded Object Detectors

University Of California-San Diego, La Jolla CA

Investigators

Abstract

Object detection cascades are one of the most significant recent developments in computer vision. By enabling real-time object detection, they are a potentially disruptive technology, which is already making a commercial impact in industries as diverse as digital photography, automotive, surveillance, personal identification, and traffic safety, among others. However, this disruptive potential is currently stifled by the substantial complexity of training detector cascades. In practice, this complexity limits the application of the cascaded architecture to a small set of domains (most notably face detection) which have been heavily researched by the academic community and for which detectors are publicly available. This project aims to eliminate the complexity hurdle, by laying the theoretical and algorithmic foundations for the fully-automated, low-complexity, design of optimal detection cascades, which guarantee high detection-rate while minimizing false-positive rate and detection complexity. In particular, the project addresses major current roadblocks in architecture design, detector design, and training complexity, through novel contributions in cost sensitive boosting, weak learners, and optimal cascade design algorithms. All contributions will be evaluated in the context of an effort to deploy real-time animal detectors in some of the most popular wild-life attractions of San Diego. This also provides an exciting and unusual opportunity for the involvement of undergraduates in research. More information on the project can be found at http://www.svcl.ucsd.edu