Concetti Chiave
AUEditNet achieves accurate manipulation of facial action unit intensities in high-resolution synthetic face images, without requiring retraining or extra estimators, by leveraging a dual-branch architecture that implicitly disentangles facial attributes and identity even with limited subject data.
Sintesi
The paper introduces AUEditNet, a method for accurately manipulating the intensities of 12 facial action units (AUs) in high-resolution synthetic face images. The key highlights are:
AUEditNet achieves impressive AU intensity manipulation performance, trained effectively with only 18 subjects, by utilizing a dual-branch architecture that implicitly disentangles facial attributes and identity without additional loss functions or large batch sizes.
The method allows conditioning the manipulation on either intensity values or target images, eliminating the need for constructing AU combinations for specific facial expression synthesis.
AUEditNet outperforms state-of-the-art AU intensity estimation and editing methods in terms of manipulation accuracy, identity preservation, and image similarity, even when evaluated on datasets with limited subject counts.
The method demonstrates the capability to transfer fine-grained facial expressions from target images without retraining the network.
Extensive experiments, including ablation studies, validate the effectiveness of AUEditNet's design choices in achieving accurate AU intensity manipulation despite the dataset's limited subject count.
Statistiche
The paper does not provide specific numerical data or statistics to support the key logics. The evaluation is primarily based on qualitative comparisons and quantitative metrics such as Intra-Class Correlation (ICC), Mean Squared Error (MSE), identity preservation, and image similarity.