DAAL: Deep activation-based attribute learning for action recognition in depth videos

In this paper, we propose a joint semantic preserving action attribute learning framework for action recognition from depth videos, which is built on multi-stream deep neural networks. More specifically, this paper describes the idea to explore action attributes learned from deep activations. Multiple stream deep neural networks rather than conventional hand-crafted low-level features are employed to learn the deep activations. An undirected graph is utilized to model the complex semantics among action attributes and is integrated into our proposed joint action attribute learning algorithm. Experiments on several public datasets for action recognition demonstrate that 1) the deep activations achieve the state-of-the-art discriminative performance as feature vectors and 2) the attribute learner can produce generic attributes, and thus obtains decent performance on zero-shot action recognition.