Action Recognition with Shot Boundary Detection and Decoded iDT Features

We report our results on the THUMOS Challenge 2014 Ac- tion Recognition Task. Given an untrimmed video as input, our method recognizes its major action category in four main steps: (1) detect shot boundaries within the video; (2) decode pre-computed indices of iDTFs within each shot; (3) encode the recovered iDTFs into shot-wise fisher vectors and compute the shot-to-category classification scores; (4) sum- marize shot-wise classification scores into a video-wise classification score. We report results on four different strategies that are used for summa- rizing the video-wise classification scores.