MULTI-TEACHER KNOWLEDGE DISTILLATION FOR COMPRESSED VIDEO ACTION RECOGNITION ON DEEP NEURAL NETWORKS

Recently, convolutional neural networks (CNNs) have seen great progress in classifying images. Action recognition is different from still image classification; video data contains temporal information that plays an important role in video understanding. Currently, most CNN-based approaches for action recognition have excessive computational costs, with an explosion of parameters and computation time. The currently most efficient method trains a deep network directly on compressed video containing the motion information. However, this method has a large number of parameters. We propose a multi-teacher knowledge distillation framework for compressed video action recognition to compress this model. With this framework, the model is compressed by transferring the knowledge from multiple teachers to a single small student model. With multi-teacher knowledge distillation, students learn better than with single-teacher knowledge distillation. Experiments show that we can reach a 2.4x compression rate in a number of parameters and a 1.2x computation reduction with 1.79% loss of accuracy on the UCF-101 dataset and 0.35% loss of accuracy on the HMDB51 dataset.