Core Concepts
Facial Action Units (AU) detection is enhanced by the innovative PETL paradigm, MoKE collaboration mechanism, and MDWA-Loss in AUFormer.
Abstract
AUFormer introduces a novel approach to Facial Action Unit (AU) detection by leveraging Parameter-Efficient Transfer Learning (PETL), a Mixture-of-Knowledge Expert (MoKE) collaboration mechanism, and a Margin-truncated Difficulty-aware Weighted Asymmetric Loss (MDWA-Loss). The method aims to address overfitting issues and improve AU detection performance without relying on additional relevant data. By integrating personalized multi-scale and correlation knowledge specific to each AU, AUFormer achieves state-of-the-art results across various domains and datasets. The collaborative approach between MoKEs and the tailored loss function contribute to the model's robustness and generalization abilities.
Stats
Existing methods suffer from overfitting due to large learnable parameters on scarce AU-annotated datasets.
PETL provides a promising paradigm for efficient fine-tuning of pre-trained models.
MoKE collaboration mechanism integrates personalized multi-scale and correlation knowledge for improved AU detection.
MDWA-Loss focuses on activated AUs, differentiates difficulty levels of unactivated AUs, and discards mislabeled samples.
Quotes
"Parameter-Efficient Transfer Learning presents a promising strategy to alleviate overfitting in fully fine-tuned models."
"MoKE collaboration mechanism efficiently leverages pre-trained Vision Transformer for AU detection."
"MDWA-Loss encourages model focus on valuable information by discarding potentially mislabeled samples."