E3D: A Scaled-up Framework for Human Action Recognition

The research represents an E3D architecture created by upscaling 2D EfficientNet architecture into a three-dimensional one. E3D is a simple yet productive network that utilises a custom and novel Depthwise 3D convolution layer on which the entire algorithm depends. A depthwise convolution layer is devised for simultaneously understanding the spatial and temporal information for a large-scale processed frame sequence dataset by leveraging and scaling EfficientNetB0. This homogeneous architecture summarises the 101 distinct actions from two publicly available action classification datasets, HMDB-51 and UCF-101. The manuscript further explains the functionality of E3D and how its trained features outshine the other state-of-the-art classification benchmark in terms of performance, efficiency and deployment characteristics by achieving 98.74% and 86.49% accuracy on the UCF-101 dataset and HMDB-51 dataset, respectively; with this, the proposed architecture is computationally cheaper than the other action recognition architectures.