2D CNN vs 3D CNN: An Empirical Study on Deep Learning-based Facial Emotion Recognition

Human emotion detection is a significant challenge in computer-based, automatic emotion detection. Deep learning techniques have been employed to extract facial features from videos, utilizing both visual and auditory input. However, the effectiveness of deep learning methods, specifically using a lightweight model, to exclusively extract facial features from videos has not been extensively explored. Consequently, this study aimed to compare the performance of two distinct 3-dimensional convolutional neural networks (CNNs) and 2-dimensional CNNs in extracting emotion from visual input in videos. Two benchmark databases, namely the Acted Facial Expressions in the Wild database and the eNTERFACE’05 Audio-Visual Emotion Dataset, were employed for evaluation. Multiple experiments conducted during this study concluded that the 2D CNN exhibited superior performance in accurately detecting facial emotion while also utilizing fewer computational resources in comparison to the 3D CNN.