Pascal Fua, Mustafa Özuysal, Daniel Weinland
Most state-of-the-art approaches to action recognition rely on global representations either by concatenating local information in a long descriptor vector or by computing a single location independent histogram. This limits their performance in presence o ...
2010
