Hyperbolic Self-Paced Multi-Expert Network for Cross-Domain Few-Shot Facial Expression Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Cross-domain few-shot facial expression recognition (CF-FER) aims to identify novel expressions using limited data from a new domain.
• Existing CF-FER methods often struggle with imbalanced datasets and fail to capture the hierarchical nature of facial expressions in Euclidean space.
• This leads to suboptimal transferable representations.

Purpose of the Study:

• To propose a novel network, the hyperbolic self-paced multi-expert network (HSM-Net), for improved CF-FER.
• To address the limitations of Euclidean space embeddings in handling imbalanced expression categories and sample difficulties.
• To enhance the modeling of hierarchical facial expression relationships and obtain more transferable features.

Main Methods:

• Developed HSM-Net featuring multiple mixture-of-experts (MoE) layers within hyperbolic space.
• Implemented a collaborative training approach using self-distillation, where experts specialize in subsets of expression categories.
• Introduced a hyperbolic self-paced learning (HSL) strategy to adaptively train the model from easy to hard samples, mitigating data imbalance issues.

Main Results:

• HSM-Net effectively models hierarchical facial expression relationships.
• The method achieves a highly transferable feature space, outperforming existing state-of-the-art approaches.
• Experiments on both in-the-lab and in-the-wild datasets validate the proposed method's superiority.

Conclusions:

• The proposed HSM-Net offers a significant advancement in cross-domain few-shot facial expression recognition.
• By leveraging hyperbolic geometry and self-paced learning, the network effectively handles data imbalances and enhances feature transferability.
• The method demonstrates strong performance on complex facial expression recognition tasks.