Oriented Classroom Instructional Behavior Recognition Benchmark

Classroom instruction behavior analysis is one of the effective methods to optimize methods and designs of classroom teaching, improving the quality and efficiency of classroom instruction, and enriching the practical knowledge of teachers. With the continuous breakthroughs in artificial intelligence technology, there is a growing body of research on instructional behavior analysis. Instruction behavior recognition algorithms face great challenges due to the lack of appropriate datasets as data support. Therefore, in this paper, we propose a dataset CIBR for classroom instructional behavior recognition. The dataset is based on real classroom teaching videos from primary and secondary schools, covering 14 instructional behaviors, with a total of 2, 380 video samples. We evaluated our CIBR dataset on five commonly used behavior recognition models, namely 3D-RESnet 50, I3D, R (2 + 1) D-RGB, S3D and Timesformer, and compared it with the mainstream behavior recognition datasets UCF101 and HMDB51. The results show that the CIBR dataset is reliable, feasible, and effective. The creation of the CIBR dataset provides a data base for AI techniques to automatically identify and analyze the behavior of teachers and students in primary and secondary schools.