TAR: A Dataset of Teacher-Teaching Action Recognition

With the rise of artificial intelligence in education, the utilization of video understanding technology to identify teachers' teaching actions is becoming increasingly important for improving the quality of teaching and accelerating the process of educational informatization. However, the lack of a dedicated dataset for teacher action recognition has been a hindrance to the development of intelligent teaching. In this paper, we construct a dataset of teacher-teaching action recognition (TAR). The dataset collected classroom videos of 25 teachers in eight classrooms of vocational and technical schools, from which a total of 13,288 valid video samples were obtained after processing. The dataset contains 8 behavior categories, including 5 categories of normative teaching behaviors and 3 categories of misbehaviors. To ensure the authenticity and representativeness of the data, all actions were collected from actual classroom videos. We then tested mainstream methods on the newly constructed dataset and proposed a Cross-channel Non-local (CNL) module based on SlowFast to capture long-range spatiotemporal dependencies. The experimental results demonstrated that the proposed method significantly outperformed other existing methods.