Learning Meta Face Recognition in Unseen Domains (2003.07733v2)

Abstract: Face recognition systems are usually faced with unseen domains in real-world applications and show unsatisfactory performance due to their poor generalization. For example, a well-trained model on webface data cannot deal with the ID vs. Spot task in surveillance scenario. In this paper, we aim to learn a generalized model that can directly handle new unseen domains without any model updating. To this end, we propose a novel face recognition method via meta-learning named Meta Face Recognition (MFR). MFR synthesizes the source/target domain shift with a meta-optimization objective, which requires the model to learn effective representations not only on synthesized source domains but also on synthesized target domains. Specifically, we build domain-shift batches through a domain-level sampling strategy and get back-propagated gradients/meta-gradients on synthesized source/target domains by optimizing multi-domain distributions. The gradients and meta-gradients are further combined to update the model to improve generalization. Besides, we propose two benchmarks for generalized face recognition evaluation. Experiments on our benchmarks validate the generalization of our method compared to several baselines and other state-of-the-arts. The proposed benchmarks will be available at https://github.com/cleardusk/MFR.

• The paper presents a novel meta-learning approach that enhances domain generalization in face recognition without extra fine-tuning.
• It employs a meta-optimization strategy with simulated domain shifts and batch sampling to robustly test models on newly introduced benchmarks GFR-R and GFR-V.
• Empirical results demonstrate improved verification rates and ranking accuracy, outperforming traditional models in real-world scenarios with unseen conditions.

Learning Meta Face Recognition in Unseen Domains

The paper "Learning Meta Face Recognition in Unseen Domains" presents a novel approach to enhancing the generalization capability of face recognition systems. Traditional face recognition models often falter when applied to domains outside of their training set due to domain-specific biases. This research proposes a Meta Face Recognition (MFR) methodology designed to address this limitation by improving domain generalization without requiring additional fine-tuning or domain-specific adjustments after initial deployment.

This method leverages a meta-learning framework that synthesizes domain shifts through a meta-optimization objective. This objective compels models to develop effective face representations capable of operating across domain boundaries. The proposed model is not only evaluated on synthesized source domains but also rigorously tested on synthesized target domains through a process of domain-level sampling. Specifically, this involves constructing batches that simulate domain shifts and integrating gradients from multiple domain distributions to refine the model. With each iteration, the network learns transferable knowledge across domains, aimed at effective deployment in varied unseen environments.

The research introduces two novel benchmarks, GFR-R and GFR-V, for testing generalized face recognition capabilities across racial and facial variability domains. GFR-R examines cross-racial accuracy by training models on certain racial domains and testing them on others. In contrast, GFR-V evaluates the model's performance on more diverse facial conditions like varying ages, lighting conditions, and occlusions. These benchmarks ensure a comprehensive analysis of the MFR's performance relative to existing baseline models and other current state-of-the-art methodologies.

Empirical results demonstrate that the MFR can outperform baseline models and domain-specific fine-tuning methods in several complex evaluation scenarios. Notable improvement in verification rates and ranking accuracy are achieved across various benchmarks, underlining its potential for robust application in real-world scenarios where deployed models encounter unfamiliar conditions.

The implications of this research extend beyond face recognition. By narrowing the gap in performance between training and unseen domains, the methodology could foster advancements in several AI applications requiring robust, domain-agnostic recognition capabilities. Future developments could explore integrating this framework into other AI systems, potentially leading to more reliable machine learning models in fields where data variability and domain shifts pose significant challenges.