| Scaling Rapid Object Detection by Crowdsourcing | ||||||
 |
||||||
| Millions of users upload images and videos to Facebook, Flickr, or YouTube every day. In the process, they unveil valuable insights about their interests, favorite brands, stores, hobbies, and entertainment. This data could be used to improve user satisfaction with the on-line experience, in several ways. For example, businesses could find new costumers (or promote their products) through coupons and special offers that explicitly address user needs (targeted advertising); social networks could be expanded through friendship suggestions based on shared interest; and users could receive the latest news about their favorite brands, places, events, or subjects. . In many cases, the information of interest is only available through the images and videos themselves. Ideally, the image could be analyzed automatically, the objects in it detected immediately, and tags produced for future use. The process should be scalable, since the average person knows about 70,000 objects, and there are billions of online images. For video, the quantities of data are taxing even for large datacenters. Overall, there is a need for highly accurate object detection algorithms of very low complexity. The goal of this project is to design the algorithms needed to scale real-time object detection to thousands of objects. The idea is to develop tools for fully automatic 1) example collection and 2) detector training. With these tools the problem can then be solved by crowd-sourcing. | ||||||
| Examples of recent detectors: | ||||||
 |
||||||
| Video demos: real-time object detection on cell phone | ||||||
| Download | ||||||
| Download | ||||||
| Related Publications: | ||||||
|
Copyright
@ 2009 www.svcl.ucsd.edu