Temporal-spatial Feature Fusion for Few-shot Skeleton-based Action Recognition

Recognizing new action categories from a few reference samples is an encouraging research field because the cost of labeling data is expensive. This work presents a method for few-shot (or one-shot) skeleton-based action recognition by fusing temporal and spatial features of actions. Trajectory primitives are proposed to characterize the temporal features, which can be obtained by segmenting and clustering the trajectories of joints. After that, we modify the original dynamic time warping (DTW) algorithm and use it to measure the similarity between trajectory primitive sequences. Besides, we compute the joint angles as spatial feature vectors. Support vector machines (SVM) are used to classify the joint angle vectors. In this way, the temporal distance matrix can be calculated by modified DTW, and the spatial distance matrix can be obtained by trained SVM. Finally, we fuse temporal and spatial distance matrices by adjusting a parameter to improve recognition accuracy. Furthermore, extensive experiments are conducted on three small-scale datasets to verify the effectiveness of our proposed method.