With the ubiquity of videos on the internet, the development of algorithms that can analyze, summarize, and classify their content is an active field of research. While Convolutional Neural Networks (CNNs, http://goo.gl/PqYGzp) are an effective class of models for understanding image content, their development and use for video classification has been limited by the lack of video datasets that match the scale and variety of existing image datasets.
With Large-scale Video Classification with Convolutional Neural Networks (http://goo.gl/6sOs3C), a paper to be presented at the 2014 Computer Vision and Pattern Recognition conference (http://goo.gl/vY2v9k), researchers from Google and Stanford University collaborate to study the performance of CNNs for large-scale video classification using Sports-1M, a new dataset consisting of ~1 million YouTube videos belonging to a taxonomy of 487 classes of sports.
Utilizing the information present in single, static frames as well as the complex temporal evolution present in video, the CNN learns features on full frame, low-resolution context streams along with centrally cropped high-resolution fovea streams. In doing so, the research shows that CNN architectures are capable of learning powerful features from video data, even if the provided description does not match the content, or there is significant variation on the frame level.
To support future work in this area, the Sports-1M dataset, consisting of 1,133,158 video URLs which have been annotated automatically with labels, has been made available to the research community at http://goo.gl/8imYDN. To see an example of per-frame classification results overlaid on top of a video, watch the video below.
- ClariDevOps, present
- Wepollscofounder, 2011 - 2013
- SpyFuSoftware Engineer, 2007 - 2011
- Microchip, Velocityscape, SpyFu
- ASUComputer Systems Engineering
Changing Costs for Americans from 2005 to 2014(relative to 23% overall i...
Regina's interactive graph and data of "Changing Costs for Americans from 2005 to 2014(relative to 23% overall increase -- inflation)" is a
Unsupervised learning, attention, and other mysteries - O'Reilly Radar
Get notified when our free report Future of Machine Intelligence: Perspectives from Leading Practitioners is available for download. The fol
whackashoe/Human_Level_Control_through_Deep_Reinforcement_Learning · GitHub
Human_Level_Control_through_Deep_Reinforcement_Learning - Deepmind's Deep Q Learning on Atari
The transformative potential of self-driving electric cars
They're more than substitutes for gas cars. They are game changers. Game changers, I tell you!
Welcome To The Unicorn Club, 2015: Learning From Billion-Dollar Companies
It’s been over a year since we wrote our original post sharing our analysis of the last decade’s most successful U.S.-based, venture-backed
Students - Guide to Technical Development - Google Careers
Guide for Technical Development. Having a solid foundation in Computer Science is important to become a successful Software Engineer. This g