Temporal-Viewpoint Transportation Plan for Skeletal Few-Shot Action Recognition

We propose a Few-shot Learning pipeline for 3D skeleton-based action recognition by Joint tEmporal and cAmera viewpoiNt alIgnmEnt (JEANIE). To factor out misalignment between query support sequences of body joints, we an advanced variant Dynamic Time Warping which jointly models each smooth path the frames to achieve simultaneously best alignment in temporal simulated camera viewpoint spaces end-to-end learning under limited few-shot training data. Sequences are encoded with block encoder based on Simple Spectral Graph Convolution, lightweight linear Neural Network backbone. also include setting transformer. Finally, similarity-based loss encourages same class while preventing unrelated sequences. show state-of-the-art results NTU-60, NTU-120, Kinetics-skeleton UWA3D Multiview Activity II.