Learning Body Language and Facial Expressions as Better Features in Semi-supervised Important People Detection

Detecting the most important individual in a group photograph, considering the overall context of the image, is a challenging problem in computer vision. Several prior research efforts have explored this task using various methodologies. Notably, deep learning-based approaches have demonstrated promising results in addressing this problem. However, a key limitation in this domain is the scarcity of annotated datasets, which significantly hinders the potential of deep learning models. Unlike traditional entity recognition tasks, which benefit from abundant labeled data, important people detection suffers from a lack of well-annotated datasets, making supervised learning approaches less effective. Furthermore, manual annotation of such datasets is not only resource-intensive but also time-consuming and difficult to scale. One of the most effective strategies to mitigate the issue of limited labeled data is to leverage semi-supervised deep learning frameworks. These approaches typically involve training a model on a small labeled dataset, using the trained model to generate pseudo-labels for a larger unlabeled dataset, and iteratively refining the model with both labeled and pseudo-labeled data. Such methodologies have been shown to significantly enhance performance in low-data scenarios. In this study, we build upon an existing state-of-the-art semi-supervised learning pipeline for important people detection [4]. We propose a novel Feature Extraction Module, which, when integrated into the existing framework, leads to substantial performance improvements. Through extensive experimentation, we demonstrate that our proposed approach enhances the overall effectiveness of the semi-supervised learning pipeline, thereby contributing to more accurate and robust important people detection in still images.