BENet: A lightweight bottom-up framework for context-aware emotion recognition
Résumé
Emotion recognition from images is a challenging task. The latest and most common approach to solve this problem is to fuse information from different contexts, such as person-centric features, scene features, object features, interactions features and so on. This requires specialized pre-trained models, and multiple pre-processing steps, resulting in long and complex frameworks, not always practicable in real time scenario with limited resources. Moreover, these methods do not deal with person detection, and treat each subject sequentially, which is even slower for scenes with many people. Therefore, we propose a new approach, based on a single end-to-end trainable architecture that can both detect and process all subjects simultaneously by creating emotion maps. We also introduce a new multitask training protocol which enhances the model predictions. Finally, we present a new baseline for emotion recognition on EMOTIC dataset, which considers the detection of the agents. Our code is available at https://github.com/TristanCladiere/BENet.git.
Domaines
Informatique [cs]
Origine : Fichiers produits par l'(les) auteur(s)