Semantic facial attribute editing using pre-trained Generative Adversarial Networks (GANs) has attracted a great deal of attention and effort from researchers in recent years. Due to the high quality of face images generated by StyleGANs, much work has focused on the StyleGANs’ latent space and the proposed methods for facial image editing. Although these methods have achieved satisfying results for manipulating user-intended attributes, they have not fulfilled the goal of preserving the identity, which is an important challenge. We present ID-Style, a new architecture capable of addressing the problem of identity loss during attribute manipulation. The key components of ID-Style include a Learnable Global Direction (LGD) module, which finds a shared and semi-sparse direction for each attribute, and an Instance-Aware Intensity Predictor (IAIP) network, which finetunes the global direction according to the input instance. Furthermore, we introduce two losses during training to enforce the LGD and IAIP to find semi-sparse semantic directions that preserve the identity of the input instance. Despite reducing the size of the network by roughly 95% as compared to similar state-of-the-art works, ID-Style outperforms baselines by 10% and 7% in identity preserving metric (FRS) and average accuracy of manipulation (mACC), respectively.

Monocular depth estimation (MDE) is crucial in a wide range of applications, including robotics, autonomous driving and virtual reality. Self-supervised monocular depth estimation has emerged as a promising MDE approach without requiring hard-to-obtain depth labels during training, and multi-scale photometric loss is widely used for self-supervised monocular depth estimation as the self-supervised signal. However, multi-photometric loss is a weak training signal and might disturb the good intermediate features representation. In this paper, we propose a successive depth map self-distillation(SDM-SD) loss, which combines with the single-scale photometric loss to replace the multi-scale photometric loss. Moreover, considering that multi-stage feature representations are essential for dense prediction tasks such as depth estimation, we also propose a dense skip connection, which can efficiently fuse the intermediate features of the encoder and fully utilize them in each stage of the decoder in our encoder–decoder architecture. By applying successive depth map self-distillation loss and dense skip connection, our proposed method can achieve state-of-the-art performance on the KITTI benchmark, and exhibit the best generalization ability on the challenging indoor dataset NYUv2 dataset.

This paper tackles the challenge of point-supervised temporal action detection, wherein only a single frame is annotated for each action instance in the training set. Most of the current methods, hindered by the sparse nature of annotated points, struggle to effectively represent the continuous structure of actions or the inherent temporal and semantic dependencies within action instances. Consequently, these methods frequently learn merely the most distinctive segments of actions, leading to the creation of incomplete action proposals. This paper proposes POTLoc, a Pseudo-label Oriented Transformer for weakly-supervised Action Localization utilizing only point-level annotation. POTLoc is designed to identify and track continuous action structures via a self-training strategy. The base model begins by generating action proposals solely with point-level supervision. These proposals undergo refinement and regression to enhance the precision of the estimated action boundaries, which subsequently results in the production of ‘pseudo-labels’ to serve as supplementary supervisory signals. The architecture of the model integrates a transformer with a temporal feature pyramid to capture video snippet dependencies and model actions of varying duration. The pseudo-labels, providing information about the coarse locations and boundaries of actions, assist in guiding the transformer for enhanced learning of action dynamics. POTLoc outperforms the state-of-the-art point-supervised methods on THUMOS’14 and ActivityNet-v1.2 datasets.

Image super resolution involves enhancing the spatial resolution of low-quality images and improving their visual quality. As in many real-life situations, the image degradation process is unknown, performing the task of image super resolution in a blind manner is of paramount importance. Deep neural networks provide high performances for the task of blind image super resolution, in view of their end-to-end learning capability between the low-resolution images and their ground truth versions. Generally speaking, deep blind image super resolution networks initially estimate the parameters of the image degradation process, such as blurring kernel, and then use them for super-resolving the low-resolution images. In this paper, we develop a novel deep learning-based scheme for the task of blind image super resolution, in which the idea of leveraging the hybrid representations is utilized. Specifically, we employ the deterministic and stochastic representations of the blurring kernel parameters to train a deep blind super resolution network in an effective manner. The results of extensive experiments prove the effectiveness of various ideas used in the development of the proposed deep blind image super resolution network.

Synthesizing complex images from text presents challenging. Compared to autoregressive and diffusion model-based methods, Generative Adversarial Network-based methods have significant advantages in terms of computational cost and generation efficiency yet remain two limitations: first, these methods often refine all features output from the previous stage indiscriminately, without considering these features are initialized gradually during the generation process; second, the sparse semantic constraints provided by the text description are typically ineffective for refining fine-grained features. These issues complicate the balance between generation quality, computational cost and inference speed. To address these issues, we propose a Multi-granularity Feature Aware Enhancement GAN (MFAE-GAN), which allows the refinement process to match the order of different granularity features being initialized. Specifically, MFAE-GAN (1) samples category-related coarse-grained features and instance-level detail-related fine-grained features at different generation stages based on different attention mechanisms in Coarse-grained Feature Enhancement (CFE) and Fine-grained Feature Enhancement (FFE) to guide the generation process spatially, (2) provides denser semantic constraints than textual semantic information through Multi-granularity Features Adaptive Batch Normalization (MFA-BN) in the process of refining fine-grained features, and (3) adopts a Global Semantics Preservation (GSP) to avoid the loss of global semantics when sampling features continuously. Extensive experimental results demonstrate that our MFAE-GAN is competitive in terms of both image generation quality and efficiency.

To retrieve a video via a multimedia search engine, a textual query is usually created by the user and then used to perform the search. Recent state-of-the-art cross-modal retrieval methods learn a joint text–video embedding space by using contrastive loss functions, which maximize the similarity of positive pairs while decreasing that of the negative pairs. Although the choice of these pairs is fundamental for the construction of the joint embedding space, the selection procedure is usually driven by the relationships found within the dataset: a positive pair is commonly formed by a video and its own caption, whereas unrelated video-caption pairs represent the negative ones. We hypothesize that this choice results in a retrieval system with limited semantics understanding, as the standard training procedure requires the system to discriminate between groundtruth and negative even though there is no difference in their semantics. Therefore, differently from the previous approaches, in this paper we propose a novel strategy for the selection of both positive and negative pairs which takes into account both the annotations and the semantic contents of the captions. By doing so, the selected negatives do not share semantic concepts with the positive pair anymore, and it is also possible to discover new positives within the dataset. Based on our hypothesis, we provide a novel design of two popular contrastive loss functions, and explore their effectiveness on four heterogeneous state-of-the-art approaches. The extensive experimental analysis conducted on four datasets, EPIC-Kitchens-100, MSR-VTT, MSVD, and Charades, validates the effectiveness of the proposed strategy, observing, e.g., more than +20% nDCG on EPIC-Kitchens-100. Furthermore, these results are corroborated with qualitative evidence both supporting our hypothesis and explaining why the proposed strategy effectively overcomes it.

Self-supervised learning methods are increasingly important for monocular depth estimation since they do not require ground-truth data during training. Although existing methods have achieved great success for better monocular depth estimation based on Convolutional Neural Networks (CNNs), the limited receptive field of CNNs usually is insufficient to effectively model the global information, e.g., relationship between foreground and background or relationship among objects, which are crucial for accurately capturing scene structure. Recently, some studies based on Transformers have attracted significant interest in computer vision. However, duo to the lack of spatial locality bias, they may fail to model the local information, e.g., fine-grained details with an image. To tackle these issues, we propose a novel self-supervised learning framework by incorporating the advantages of both the CNNs and Transformers so as to model the complete contextual information for high-quality monocular depth estimation. Specifically, the proposed method mainly includes two branches, where the Transformer branch is considered to capture the global information while the Convolution branch is exploited to preserve the local information. We also design a rectangle convolution module with pyramid structure to perceive the semi-global information, e.g. thin objects, along the horizontal and vertical directions within an image. Moreover, we propose a shape refinement module by learning the affinity matrix between pixel and its neighborhood to obtain accurate geometrical structure of scenes. Extensive experiments evaluated on KITTI, Cityscapes and Make3D dataset demonstrate that the proposed method achieves the competitive result compared with the state-of-the-art self-supervised monocular depth estimation methods and shows good cross-dataset generalization ability.

Object Re-Identification (Re-ID) aims to identify and retrieve specific objects from images captured at different places and times. Recently, object Re-ID has achieved great success with the advances of Vision Transformers (ViT). However, the effects of the global–local relation have not been fully explored in Transformers for object Re-ID. In this work, we first explore the influence of global and local features of ViT and then further propose a novel Global–Local Transformer (GLTrans) for high-performance object Re-ID. We find that the features from last few layers of ViT already have a strong representational ability, and the global and local information can mutually enhance each other. Based on this fact, we propose a Global Aggregation Encoder (GAE) to utilize the class tokens of the last few Transformer layers and learn comprehensive global features effectively. Meanwhile, we propose the Local Multi-layer Fusion (LMF) which leverages both the global cues from GAE and multi-layer patch tokens to explore the discriminative local representations. Extensive experiments demonstrate that our proposed method achieves superior performance on four object Re-ID benchmarks. The code is available at https://github.com/AWangYQ/GLTrans.

Automatic font generation is of great benefit to improving the efficiency of font designers. Few-shot font generation aims to generate new fonts from a few reference samples, and has recently attracted a lot of attention from researchers. This is valuable but challenging, especially for ideograms with high diversity and complex structures. Existing models based on convolutional neural networks (CNNs) struggle to generate glyphs with accurate font style and stroke details in the few-shot setting. This paper proposes the TransFont, exploiting the long-range dependency modeling ability of the Vision Transformer (ViT) for few-shot font generation. For the first time, we empirically show that the ViT is better at glyph image generation than CNNs. Furthermore, based on the observation of the high redundancy in the glyph feature map, we introduce the glyph self-attention module for mitigating the quadratic computational and memory complexity of the pixel-level glyph image generation, along with several new techniques, i.e., multi-head multiple sampling, yz axis convolution, and approximate relative position bias. Extensive experiments on two Chinese font libraries show the superiority of our method over existing CNN-based font generation models, the proposed TransFont generates glyph images with more accurate font style and stroke details.

The challenge in multi-focus image fusion tasks lies in accurately preserving the complementary information from the source images in the fused image. However, existing datasets often lack ground truth images, making it difficult for some full-reference loss functions (such as SSIM) to effectively participate in model training, thereby further affecting the performance of retaining source image details. To address this issue, this paper proposes an unsupervised dual-channel dense convolutional method, DCD, for multi-focus image fusion. DCD designs Patch processing blocks specifically for the fusion task, which segment the source image pairs into equally sized patches and evaluate their information to obtain a reconstructed image and a set of adaptive weight coefficients. The reconstructed image is used as the reference image, enabling unsupervised methods to utilize full-reference loss functions in training and overcoming the challenge of lacking labeled data in the training set. Furthermore, considering that the human visual system (HVS) is more sensitive to brightness than color, DCD trains the dual-channel network using both RGB images and their luminance components. This allows the network to focus more on the brightness information while preserving the color and gradient details of the source images, resulting in fused images that are more compatible with the HVS. The adaptive weight coefficients obtained through the Patch processing blocks are also used to determine the degree of preservation of the brightness information in the source images. Finally, comparative experiments on different datasets also demonstrate the superior performance of DCD in terms of fused image quality compared to other methods.