With growth of the Internet, vast amounts of visual data are being produced. These data are from mobile phones, laptops and digital cameras. Moreover, network cameras are deployed for various purposes such as surveillance, traffic monitoring, and sightseeing. The visual data contains valuable information. However, without processing, the information will be lost. Analyzing large-scale visual data requires significant amounts of computing resources and costs. One of the solutions is cloud computing. However, it may still be expensive when processing the data. Some cloud vendors (such as Amazon) introduce spot instances that use spare instances with dynamic pricing. The spot instances offer the same performance as on-demand instances but spot instances may be terminated at short notice. As a result, processing programs may not finish when using spot instances. This dissertation introduces a large-scale image processing system that utilizes spot instances to reduce computing cost. This system uses a check-pointing method to store progress so that processing can resume later if the spot instances are terminated.

One of the possible applications that utilize deployed network cameras is improving public safety. Although the cameras are not deployed for surveillance purposes, the cameras can be utilized to increase public safety by properly integrating to currunt surveillance systems. Suspicious activities may be monitored in real-time and coverage can be increased along with the closed-circuit televisions (CCTV) deployed by law enforcement. Integrating public cameras into a surveillance system has many challenges such as inaccurate locations, diverse sources, and different methods to access the visual data. This dissertation presents how to discover public cameras from hetrogeneous sources and find the accurate locations and orientations of the cameras.