Ullas Gargi

The YouTube video recommendation system

We discuss the video recommendation system in use at YouTube, the worlds most popular online video community. The system recommends personalized sets of videos to users based on their activity on the site. We discuss some of the unique challenges that the system faces and how we address them. In addition, we provide details on the experimentation and evaluation framework used to test and tune new algorithms. We also present some of the findings from these experiments.

Detecting highlights in sports videos: Cricket as a test case

In this paper, we propose a novel approach for detecting highlights in sports videos. The videos are temporally decomposed into a series of events based on an unsupervised event discovery and detection framework. The framework solely depends on easy-to-extract low-level visual features such as color histogram (CH) or histogram of oriented gradients (HOG), which can potentially be generalized to different sports. The unigram and bigram statistics of the detected events are then used to provide a compact representation of the video. The effectiveness of the proposed representation is demonstrated on cricket video classification: Highlight vs. Non-Highlight for individual video clips (7000 training and 7000 test instances). We achieve a low equal error rate of 12.1% using event statistics based on CH and HOG features.

Solving the label resolution problem in supervised video content classification

Classifying video based on its content often requires learning a classifier from labeled samples. Sometimes the semantic labels available for a video refer to the class or category of the video as a whole; whereas the discriminative features that result in that categorization may only occur over a temporal subset of the video and may occur anywhere, leading to sub-optimal performance. We therefore need to simultaneously learn discriminative features and their temporal support while remaining independent of position within the video. We propose a solution to this label-resolution problem using a wavelet decomposition of frame-level feature time-series followed by learning discrimin ative extrema values over multiple time scales. We apply this approach to automatically detecting the presence of adult content in online videos with a resulting equal-error rate around 5%.

QuickSuggest: character prediction on web appliances

As traditional media and information devices integrate with the web, they must suddenly support a vastly larger database of relevant items. Many devices use remote controls with on-screen keyboards which are not well suited for text entry but are difficult to displace. We introduce a text entry method which significantly improves text entry speed for on-screen keyboards using the same simple Up/Down/Left/Right/Enter interface common to remote controls and gaming devices used to enter text. The paper describes QuickSuggests novel adaptive user interface, demonstrates quantitative improvements from simulation results on millions of user queries and shows ease of use and efficiency with no learning curve in user experiments.