Core Concepts
Facial expression analysis can automate user preference annotation for image generation tasks, improving scalability and efficiency.
Abstract
1. Abstract:
Proposes automatic annotation of user preferences from facial expressions to enhance text-to-image generative models.
Introduces the FERGI dataset correlating facial action units (AUs) with user evaluations of generated images.
2. Introduction:
Discusses limitations in human feedback collection for model fine-tuning due to manual annotation reliance.
Presents a method to automatically annotate user preferences using facial expression reactions.
3. Related Work:
Reviews various text-to-image generation models and evaluation metrics.
Highlights the importance of training human preference scoring models based on large datasets.
4. FERGI Dataset:
Describes data collection procedure and participant details.
Explains AU model training and facial feature extraction process.
5. AU Model Training:
Details data filtering process and computation of AU activation values.
6. Experiments:
Analyzes statistical relationships between AU activation values and user evaluations.
Evaluates the performance of the AUcomb valence score in predicting image preferences independently.
7. Conclusion:
Suggests potential applications beyond text-to-image generation tasks.