In the Face Super-Resolution (FSR) task, it is important to precisely recover facial textures while maintaining facial contours for realistic high resolution faces. Although several CNN-based FSR methods have achieved great performance, they fail in restoring the facial contours due to the limitation of local convolutions. In contrast, Transformer-based methods which use self-attention as the basic component, are expert in modeling long-range dependencies between image patches. However, learning long-range dependencies often deteriorates facial textures due to the lack of locality. Therefore, a question is naturally raised: how to effectively combine the superiority of CNN and Transformer for better reconstructing faces? To address this issue, we propose an Efficient Latent Style guided Transformer-CNN framework for FSR called ELSFace , which can sufficiently integrate the advantages of CNN and Transformer. The framework consists of a Feature Preparation Stage and a Feature Carving Stage. Basic facial contours and textures are generated in the Feature Preparation Stage, and separately guided by latent styles, so that facial details are better represented in reconstruction. CNN and Transformer streams in the Feature Carving Stage are used to individually restore facial textures and facial contours, respectively in a parallel recursive way. Considering the negligence of high-frequency features when learning the long-range dependencies, we design the High-Frequency Enhancement Block (HFEB) in the Transformer stream. The Sharp Loss is also proposed for better perceptual quality in optimization. Extensive experimental results demonstrate that our ELSFace can achieve the best results among all metrics compared to the state-of-the-art CNN and Transformer-based methods on commonly used datasets and real-world tasks. Meanwhile, our ELSFace method has the least model parameters and running time. The codes are released at https://github.com/FVL2020/ELSFace .

Human Action Recognition (HAR) in night video sensing has become a crucial task for a wide range of applications (e.g., night surveillance and self-driving at night). Recently, Fully Supervised Learning (FSL) methods by training with large-scale labeled data achieve substantial performance in HAR. However, since the data captured by red–green–blue (RGB) sensor-based cameras in the night (low-light) scenes suffer from poor illumination, it leads to difficult annotations and limits the applications of night HAR by the FSL method. A potential solution for this issue is to transfer knowledge learned from normal-light videos to low-light ones using Semi-Supervised Learning (SSL) methods, such as Unsupervised Domain Adaptation (UDA) and self-training. Although, these SSL methods have shown promising results in datasets with the same or similar data distributions, they struggle with the dataset with large domain discrepancies, e.g., normal-light and low-light conditions. To address this issue, a new SSL method called V-DixMatch, which includes the pixel-level adaptation, the feature-level adaptation, and the Cross-Domain Video-based (CDV) self-training, is proposed in this work. Specifically, the pixel-level adaptation reduces the domain discrepancy in low-level features by pixel-level processing and the feature-level adaptation approximates domain discrepancy in high-level features through adversarial learning. Then, CDV self-training further provides a robust self-training strategy by designing a video-based Blended Frames Augmentation (BFA), which is a data augmentation method tailored for video-based data, and a pseudo-label collection, which improves the quality of pseudo-labels in self-training. Extensive experimental results demonstrate that V-DixMatch achieves state-of-the-art performance in the SSL low-light HAR, and is even comparable to some FSL methods.