Exploring Implicit Human Social Relationships through CNN-Based Algorithms

Sammanfattning

Computer vision is a part of Artificial Intelligence that empowers machines to perceive and comprehend data similarly to how humans do naturally. Human connections are primarily established through interactions, between individuals in society. In settings and contexts captured in images that involve beings interacting with each other can be analyzed and interpreted using computer vision technology. These interactions can vary from ones (explicit relationships) to more subtle ones (implicit relationships) which depend heavily upon the surrounding circumstances and settings. This study centers around the task of recognizing both social bonds, among people depicted in images by utilizing algorithms based on Convolutional Neural Networks (CNNs).The proposed model, referred to as the dual-glance model, the first glance focuses on detecting and localizing individuals, while the second glance enhances relationship understanding through innovations such as advanced attention mechanisms, including MultiheadAttention and Global Context Blocks. These innovations allow the model to capture both local details and broader scene context, improving accuracy in detecting social relationships.The performance of the model was evaluated using the People in Social Context (PISC) dataset, specifically focusing on two relationship classes: Friend and Commercial. The results demonstrate significant performance improvements, with the mean Average Precision (mAP) for detecting Friend relationships increasing from 49.1% to 61.6% and for Commercial relationships from 53.9% to 67.2%. These improvements emphasize the enhanced ability of the model to detect and interpret both implicit and explicit social relationships, providing valuable insights into the field of computer vision and advancing the state-of-the-art in social relationship detection.