Paincontrol: identity-preserving pain expression transfer with generative diffusion models.

Area of Science:

• Affective Computing
• Computer Vision
• Medical AI

Background:

• Automated pain detection requires large, diverse facial expression datasets.
• Acquiring real pain expression data is challenging due to privacy and ethical concerns.
• Existing methods for synthetic expression generation face limitations in accuracy and identity preservation.

Purpose of the Study:

• To explore synthetic pain expression generation for augmenting limited real pain datasets.
• To introduce PainControl, a novel method for identity-preserving pain expression transfer.
• To evaluate the effectiveness of synthetic data in improving pain detection models.

Main Methods:

• Developed PainControl, a facial landmark-guided ControlNet adaptation for expression transfer.
• Utilized text and facial landmarks to control muscle activations while preserving identity via image embeddings.
• Compared PainControl against existing methods, assessing perceptual realism, identity preservation, AU transfer accuracy, and downstream pain detection performance.

Main Results:

• PainControl generated realistic synthetic pain expressions, outperforming baselines in human evaluations.
• Challenges remain in accurately synthesizing expression intensities for key facial action units (AU4, AU43).
• Synthetic augmentation improved pain detection performance significantly in data-scarce regimes, despite not enhancing performance when augmenting larger real datasets.

Conclusions:

• PainControl demonstrates viability as a data augmentation tool for pain detection, especially when real data acquisition is difficult or ethically constrained.
• The study validates the practical use of expression transfer methods to overcome data scarcity in affective computing.
• Future research should focus on AU alignment and temporal modeling for clinical-grade synthetic pain expression generation.