25 March 2025 in IEEE Access
Remote photoplethysmography (rPPG) estimation has made considerable progress by leveraging deep learning, yet its performance remains highly susceptible to the domain shifts caused by lighting, skin tone and/or movement, particularly during inference. Additionally, continuous adaptation across multiple domains is essential due to the dynamic environmental changes such as lighting transitions or continuous motions. Domain adaptation has been widely investigated mostly focusing on classification tasks with labels. Prior arts to address domain shifts in rPPG estimation, a regression task, relies on labeled data, requires pre-training on target domains, and focuses on single-domain test-time adaptation (TTA). However, there are still remaining challenges in TTA for their applicability of rPPG estimation in real-world scenarios such as no label during inference, continuous adaptation over multiple domains, and catastrophic forgetting when re-adapting to the source domain. In this work, we recast a rPPG TTA problem as a continual learning and propose an efficient continual TTA method that mitigates significant domain shifts in target domains without labels by leveraging the non-contrastive unsupervised learning loss with selectively updating only the batch normalization layers as well as alleviates catastrophic forgetting in source domain by adopting the learning without forgetting (LwF) regularization in the frequency domain. Our method without target labels consistently yielded performance improvements in challenging continual adaptation scenarios including adapting to multiple new domain datasets over several cycles. This approach does not only mitigate catastrophic forgetting in source domain, but also ensures robust performance across different domains.
5 November 2025 in Image and Vision Computing
15 July 2025 in IEEE Access
Recent advancements in image sensor technology have introduced non-Bayer color filter arrays (CFAs), such as Quad (2 × 2), Nona (3 × 3), and Q×Q (4 × 4) patterns, which improve performance under low-light conditions compared to the traditional Bayer CFA. Although numerous deep learning models exist for Bayer demosaicing, directly applying them to non-Bayer CFAs often yields suboptimal results, despite identical input-output dimensions and differences only in pixel-level color arrangements. In this work, we identify pixel discontinuity, a structural mismatch introduced by modern non-Bayer CFAs, as the key factor behind this performance gap. Then, based on our investigation, we propose a simple yet effective solution to mitigate it: inserting a pixel unshuffle layer at the input of existing Bayer demosaicing networks. This key operation spatially restructures the input, enabling seamless adaptation to non-Bayer patterns without altering the original network architecture. Our method significantly improves demosaicing performance with almost no increase in both network size and computational cost. Moreover, it generalizes across various CFA patterns, including Quad, Nona, and Q×Q, facilitating the reuse of Bayer-trained models for a wide range of sensor designs. Experimental results on synthetic and real non-Bayer data demonstrate the practical effectiveness of our approach, removing the need for custom model design for other non-Bayer CFA types.