Bayesian 3D ConvNets for Action Recognition from Few Examples

A core challenge in action recognition from videos is obtaining sufficient training examples to train deep networks. This holds especially for action tasks from non-standard sensors such as infra-red cameras. In this work, we investigate Bayesian 3D ConvNets for action recognition when training examples are scarce. This work connects 3D ConvNets, a state-of-the-art approach for action recognition, with Bayesian networks, which have shown to be effective regularizers for deep networks. We do so by extending Bayes by Backprop to 3D ConvNets. Experimental evaluation on three small-scale action datasets from both RGB and infra-red sensors shows that Bayesian 3D ConvNets have a better test generalizing than standard 3D ConvNets. We find that, the more scarce the number of training examples per action, the better our Bayesian 3D ConvNets performs, highlighting the potential of Bayesian learning in the video domain.