Leveraging Context for Enhanced Emotion Recognition: A Study Using SSD with ResNet

The increasing interest in automated emotional state recognition systems presents significant potential for applications in human-machine interactions and monitoring. Nonetheless, conventional methods that concentrate on facial expressions and body language for emotion recognition often disregard the substantial variations stemming from cultural and situational contexts. This paper addresses the limitations of conventional emotion recognition methods, which often overlook cultural and situational context. We propose an extension to the system introduced in “Context-Based Emotion Recognition using the EMOTIC Dataset” to classify and detect 19 out of 26 emotions. Our approach leverages Single Shot Detectors (SSD) like ResNet50 and ResNet-101 as backbones. Using a subset of EMOTIC dataset consisting of 9,000 images, our model, based on SSD with ResNet-50 and a batch size of 8, we achieved an 8% higher recall than the second-best model and a 35% improvement over the worst. We also address dataset bias and propose improved annotation metrics. This research highlights the critical role of context in emotion recognition and the potential of context-aware deep learning for human-computer interaction and monitoring systems.