Abstract
The emergence of low-cost cameras with nearly professional features in the consumer market represents a new important source of video information. For example, using an increasing number of these cameras in live TV broadcastings enables obtaining varied contents without affecting the production costs. However, searching for interesting shots (e.g., a certain view of a specific car in a race) among many video sources in real-time can be difficult for a Technical Director (TD). So, TDs require a mechanism to easily and precisely represent the kind of shot they want to obtain abstracting them from the need to be aware of all the views provided by the cameras. In this paper we present our proposal to help a TD to visually define, using an interface for the definition of 3D scenes, an interesting sample view of one or more objects in the scenario. We recreate the views of the cameras in a 3D engine and apply 3D geometric computations on their virtual view, instead of analyzing the real images they provide, to enable an efficient and precise real-time selection. Specifically, our system computes a similarity measure to rank the candidate cameras. Moreover, we present a prototype of the system and an experimental evaluation that shows the interest of our proposal.
Highlights
Nowadays the consumer market offers cameras with very interesting features at low prices
As our approach depends on the location and direction of the objects and cameras in the scenario, and in the real scenario we consider in our tests that this information is updated once every second, the system has to be able to perform the processing in less than one second to provide the Technical Director (TD) with updated and accurate information
In [11] the authors emphasized that “image retrieval is only meaningful in its service to people, performance characterization must be grounded in human evaluation”
Summary
Nowadays the consumer market offers cameras with very interesting features at low prices. One important difference with existing works is that, instead of applying traditional image processing techniques over 2D projections of 3D query scenes and the candidate images, we propose obtaining high-level features (related to the semantics of the specific objects in the scene) directly from the 3D scenario. In this way, information such as the specific objects in the scene, the percentage of each object shown, the percentage of the scene filled by each object, the specific viewpoint of each object shown, etc., can be obtained precisely and in real-time
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
