Researchers have developed AffectFlow-DINO, a novel system for multi-task affect estimation from facial images. This system extends a standard architecture with a conditional rectified-flow head, allowing it to model the ambiguity in facial expressions and provide uncertainty-aware predictions. Built on a DINOv3 backbone, AffectFlow-DINO jointly estimates valence-arousal, classifies eight facial expressions, and detects twelve Action Units, significantly outperforming the official baseline for the 11th ABAW challenge. AI
IMPACT This research could improve the accuracy and robustness of AI systems analyzing human emotions and expressions.
RANK_REASON The cluster describes a new research paper detailing a novel model for affect estimation.
Read on Hugging Face Daily Papers →
- ABAW Challenge
- AffectFlow-DINO
- alphaXiv
- arXiv
- DagsHub
- DINOv3 ViT-S/16
- Hugging Face
- Salah Eddine Bekhouche
- 11th ABAW Challenge
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →