Stable Bidirectional Graph Convolutional Networks for Label-Frugal Skeleton-based Recognition
Keywords: Stable Bidirectional Graph Convolutional Networks (GCNs), Generative GCNs, Label-Efficient Training, Skeleton-based Recognition
Abstract: Skeleton-based action recognition is a major challenge in computer vision. In particular, solutions based on graph convolutional networks (GCNs) have demonstrated notable performance, but their success is reliant on the availability of large collections of hand-labeled skeleton sequences. However, in real-world applications, these sequences are often scarce, prompting the exploration of label-frugal GCN models. In this paper, we introduce a novel label-efficient GCN model for skeleton-based action recognition. As a first contribution, we devise a new acquisition function that allows us to design exemplars (a few candidate data for labeling) using an adversarial objective function that mixes representativity, diversity and uncertainty of these exemplars. As a second contribution, we make our designed GCNs bidirectional and stable, allowing them to map data from ambient to latent spaces (and vice-versa) where the inherent distribution of the learned exemplars is more easily captured. Extensive experiments conducted on two challenging skeleton-recognition datasets, show a substantial gain of our frugally designed GCNs against the related work.
Submission Number: 73
Loading