Walking and talking: A bilinear approach to multi-label action recognition

Action recognition is a fundamental problem in computer vision. However, all the current approaches pose the problem in a multi-class setting, where each actor is modeled as performing a single action at a time. In this work we pose the action recognition as a multi-label problem, i.e., an actor can be performing any plausible subset of actions. Determining which subsets of labels can co-occur is typically treated as a separate problem, typically modeled sparsely or fixed apriori to label correlation coefficients. In contrast, we formulate multi-label training and label correlation estimation as a joint max-margin bilinear classification problem. Our joint approach effectively trains discriminative bilinear classifiers that leverage label correlations. To evaluate our approach we relabeled the UCLA Courtyard dataset for the multi-label setting. We demonstrate that our joint model outperforms baselines on the same task and report state-of-the-art per-label accuracies on the dataset.