Evaluating the Performance of Mobile-Convolutional Neural Networks for Spatial and Temporal Human Action Recognition Analysis

Human action recognition is a computer vision task that identifies how a person or a group acts on a video sequence. Various methods that rely on deep-learning techniques, such as two- or three-dimensional convolutional neural networks (2D-CNNs, 3D-CNNs), recurrent neural networks (RNNs), and vision transformers (ViT), have been proposed to address this problem over the years. Motivated by the fact that most of the used CNNs in human action recognition present high complexity, and the necessity of implementations on mobile platforms that are characterized by restricted computational resources, in this article, we conduct an extensive evaluation protocol over the performance metrics of five lightweight architectures. In particular, we examine how these mobile-oriented CNNs (viz., ShuffleNet-v2, EfficientNet-b0, MobileNet-v3, and GhostNet) execute in spatial analysis compared to a recent tiny ViT, namely EVA-02-Ti, and a higher computational model, ResNet-50. Our models, previously trained on ImageNet and BU101, are measured for their classification accuracy on HMDB51, UCF101, and six classes of the NTU dataset. The average and max scores, as well as the voting approaches, are generated through three and fifteen RGB frames of each video, while two different rates for the dropout layers were assessed during the training. Last, a temporal analysis via multiple types of RNNs that employ features extracted by the trained networks is examined. Our results reveal that EfficientNet-b0 and EVA-02-Ti surpass the other mobile-CNNs, achieving comparable or superior performance to ResNet-50.