March 26:  Graham Taylor, CIMS

Learning good features for understanding video data
  

  Currently the best performing methods for recognizing human activities in video are based on engineered descriptors with explicit local geometric cues and other heuristics. I will talk about an alternative method that uses unsupervised learning to extract both perceptually relevant low-level features that are sensitive to motion and sparse mid-level features that capture longer-term temporal effects. We apply our method to recognizing actions in Hollywood movies. This is joint work with Rob Fergus, Chris Bregler and Yann LeCun.