Personalized Feature Translation for Expression Recognition: An Efficient Source-Free Domain Adaptation Method
Masoumeh Sharafi, Soufiane Belharbi, Muhammad Osama Zeeshan, Houssem Ben Salem, Ali Etemad, Alessandro Lameiras Koerich, Marco Pedersoli, Simon Bacon, Eric Granger
TL;DR
This paper introduces SFDA-PFT, a novel source-free domain adaptation method that personalizes facial expression recognition models in privacy-sensitive scenarios by translating features in latent space without generating images.
Contribution
It proposes a lightweight latent-space translation approach that adapts to neutral target data without source data or image synthesis, improving FER performance in privacy-constrained settings.
Findings
Outperforms state-of-the-art SFDA methods on multiple datasets
Reduces computation by avoiding image generation
Effectively adapts to neutral target data in privacy-sensitive scenarios
Abstract
Facial expression recognition (FER) models are widely used in video-based affective computing applications, such as human-computer interaction and healthcare monitoring. However, deep FER models often struggle with subtle expressions and high inter-subject variability, limiting performance in real-world settings. Source-free domain adaptation (SFDA) has been proposed to personalize a pretrained source model using only unlabeled target data, avoiding privacy, storage, and transmission constraints. We address a particularly challenging setting where source data is unavailable and the target data contains only neutral expressions. Existing SFDA methods are not designed for adaptation from a single target class, while generating non-neutral facial images is often unstable and expensive. To address this, we propose Source-Free Domain Adaptation with Personalized Feature Translation…
Peer Reviews
Decision·ICLR 2026 Poster
（1） Structure: The paper features a clear structure, with algorithms presented in easily understandable diagrams. The problem definition is precise, and the motivation is well-articulated. （2） Innovation: The experiments introduce a personalized translation approach within the feature space, circumventing the instability inherent in traditional image-based methods. （3） Quality: The experiment design is rigorous, validating the method's effectiveness across four distinct datasets. Detailed ablati
（1） Although comparisons are made with multiple SFDA methods, the paper does not include more recently proposed personalized approaches based on generative models or meta-learning. （2） The paper notes performance degradation on elderly subjects but does not propose adaptive strategies for age differences. Further exploration of stratified or age-aware adaptation methods is recommended.
This paper introduces a highly original and impactful approach to source-free domain adaptation for facial expression recognition. By formulating a novel feature-level translation method that operates using only neutral target data, it achieves state-of-the-art performance while being dramatically more efficient than image-based alternatives. The work is exceptionally well-supported through rigorous experiments on four diverse benchmarks and presents a practical solution to key real-world constr
1． While the proposed PFT method is well-motivated, the paper lacks a clear theoretical or intuitive explanation of why feature translation in latent space is inherently more robust than image-level translation for expression preservation. 2． The paper lacks a rigorous explanation or analysis of how the proposed losses ensure that the translator modifies only identity-related features while preserving expression-related ones. 3． The proposed PFT method relies on pre-training a feature-space tr
The key strength of the proposed PFT method lies in its innovative feature-space translation paradigm. This method avoids the need for complex image synthesis and achieves computational efficiency through lightweight parameter adaptation. It demonstrated superior performance over state-of-the-art methods across four FER benchmarks.
1 While the empirical results are strong, the paper lacks an analysis explaining the reasons why feature-space translation is more effective. 2 The experimental validation is centered primarily on FER. It would be valuable to discuss the potential of PFT for other tasks that require subject-specific adaptation, such as face recognition or person re-identification.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.