This article introduces video action recognition, covering its basic definition, downstream tasks, major algorithmic families—including CNN‑based, Vision‑Transformer, self‑supervised, and multimodal approaches—and discusses practical deployment scenarios and open challenges in the field.
This article introduces a lightweight spatiotemporal self‑attention transformer, called Patch Shift Transformer, which achieves competitive video action recognition performance on datasets such as Kinetics‑400, Sth‑v1/v2, and Diving48 without increasing computational cost or parameters, and details its design, experiments, and speed advantages.
Alibaba Cloud’s Multimedia AI team won five first‑place titles and one runner‑up across six major video‑action challenges at CVPR2021, showcasing advanced transformer‑CNN hybrids, self‑supervised initialization, and spatio‑temporal relation modeling that now power their multimedia AI cloud products.
This survey reviews video action recognition, comparing 3D convolutional networks that jointly model spatial‑temporal cues but are computationally heavy with 2D‑based approaches like TSM and TIN that embed temporal shifts efficiently, and emphasizes how large‑scale pre‑training markedly improves performance despite limited labeled data.
