We investigate architectures of discriminatively trained deep Convolutional
Networks (ConvNets) for action recognition in video. The challenge is to
capture the complementary information on appearance from still frames and
motion between frames. We also aim to generalise the best perfo