Human Action Recognition (HAR) remains a difficult task in computer vision due to the challenge of simultaneously extracting spatial and temporal information from video data. While various methods leveraging 2D and 3D convolutional neural networks (CNNs) have achieved success, the problem persists. In this study, we conduct a comparative analysis of two leading models—Inflated 3D CNN (I3D) and SlowFast networks—using the UCF-101 and JHMDB-51 datasets to evaluate their effectiveness in addressing this challenge. We developed a unified framework to standardize the evaluation process, ensuring a fair and rigorous comparison between the models. Our experimental results indicate that the I3D model outperforms SlowFast, achieving 95.4% accuracy on UCF-101 and 80.8% on JHMDB-51, compared to SlowFast’s 94.2% and 80.2%, respectively. Despite advancements in efficiency, our findings suggest that significant challenges remain in improving the accuracy of HAR. The implementation code is publicly available at https://github.com/Las-kh/actionrecognition.