ECCV 2016 Daily - Friday

In many machine-learning tasks , it is needed to represent a sequence of variable length by a fixed-length vector. For example, representing a video which is a sequence of frames, or a sentence - a sequence of words. In this paper, the team proposes the RNN Fisher Vector as an effective representation for sequences. The methodology they use is based on Fisher Vectors, where RNNs are the generative probabilistic models (instead of GMM which is used in the traditional Fisher Vector). The team shows that this representation can be computed effectively using backpropagation, and report state-of-the-art results obtained in two central but distant tasks: video action recognition and image annotation. The project is conducted by a team composed by: Guy Lev, Gil Sadeh, Benjamin Klein and Lior Wolf. Visit their poster, it’s today (Friday) at 11:00-12:30 [poster P-4A-30 ] RNN Fisher Vectors for Action Recognition and Image Annotation 22 ECCV Daily : Fr iday Presentations