RA-GCNv2 for robust skeleton-based action recognition.

What is it about?

A multi-stream graph convolutional network to explore sufficient discriminative features spreading over all skeleton joints, so that the distributed redundant representation reduces the sensitivity of the action models to non-standard skeletons.

Featured Image

Why is it important?

1. In previous work, the activation masks are obtained by a Softmax function in the activation modules, which activates only a few joints for each stream. In contrast, we propose to use a normalization activation function to expand the activated scope, thus the corresponding stream will obtain a better and more interpretable activation map. 2. Compared to previous work, we extend the original loss function with a number of additional cross-entropy regularizations on each individual network stream, so that the features can be learnt more effectively. 3. The synthetic datasets are extended by more degradation operators, where the occlusion degradation is further divided into four types, including Frame, Part, Block and Random, and two synthetic jittering datasets are newly constructed. More experiments are performed to validate the effectiveness and robustness of the proposed approach in different degradation conditions.

Perspectives