Single Image Dehazing is a challenging ill-posed problem. While Transformers excel in long-range modeling, the low-pass characteristic of their self-attention further obscures high-frequency scene details attenuated by scattering. Furthermore, haze-degraded encoder features exhibit Mean-Elevation and Variance-Compression, leading to a significant statistical discrepancy with clear decoder features that hinders effective feature fusion. To address these limitations, this paper proposes the Frequency-Domain Stepwise and Moment Adjustment Network (FDSMA-Net). First, we design the High-Frequency Disentangled Attention (HFDA) mechanism, which explicitly decouples high-frequency components via attention inversion and utilizes a Dual Dynamic Frequency Modulation mechanism to calibrate and enhance details, effectively overcoming the spectral bias of standard attention. Second, the Moment-Driven Channel Fusion Attention (MDCFA) mechanism is introduced in the skip connections of the U-Net architecture. By explicitly aligning the first-order moment and second-order moment, MDCFA dynamically corrects the distribution shift caused by haze, ensuring statistically aligned feature fusion. Enabled by frequency disentanglement and moment alignment, FDSMA-Net excels on synthetic and real benchmarks, effectively recovering scene details and enhancing visual clarity. The source code of our proposed FDSMA-Net will be released at https://github.com/SWU-CS-MediaLab/FDSMA-Net.

Depth completion reconstructs dense depth maps from sparse LiDAR measurements and synchronized RGB images. Existing methods mainly face two challenges: the texture-geometry ambiguity and the topological rigidity propagation. To overcome these limitations, we propose a unified framework of YCbCr color space and the Dual-path Attention Network (YDANet), which treats the depth completion pipeline as a Disentanglement-Regression-Propagation paradigm. In the YDANet, we first introduce the YCbCr-Guided Adaptive Fusion (YGAF) module, which explicitly disentangles structural luminance from texture-rich chrominance and exploits intrinsic illumination-depth correlations to mitigate the texture-copying artifacts common in RGB-based methods. Then, to bridge the semantic gap between the backbone and refinement heads, we propose the Dual-Path Attention (DPA) module. Functioning as a lightweight Guidance-Refiner, DPA employs Spatial Reduction Attention (SRA) for global context and CNN-CBAM for local fidelity, utilizing a channel-wise fusion mechanism to distill clean affinity priors. Additionally, we incorporate the Gated Deformable Spatial Propagation Network (GD-SPN) to overcome the topological rigidity propagation. By employing a competitive gating mechanism to ensemble heterogeneous propagation kernels and sharing learned geometric parameters across iterations, GD-SPN achieves efficient non-rigid refinement. Extensive experiments demonstrate that YDANet achieves state-of-the-art performance among all 2D methods on the NYU Depth v2 dataset and ranks first on the KITTI depth completion online leaderboard among real-time approaches (< 80 ms) at the time of submission. The source code of our proposed YDANet will be released at https://github.com/SWU-CS-MediaLab/YDANet.

Continuous Learning (CL) enables deep neural networks to adapt to evolving data while preserving performance on previously learned tasks, without access to all historical data. A significant challenge in Continuous Learning is the risk of over-fitting to limited new-class data and catastrophic forgetting of prior knowledge. Although recent methods have demonstrated that pre-trained Vision Transformers (ViTs) with prompt tuning can mitigate these issues, these methods rely on fixed, learnable prompts, which limit knowledge transfer and lead to sub-optimal performance. To address these limitations, we propose a ViT-Enhanced Prompts (VEP) framework for the tasks of continuous learning based on the pre-trained ViTs. Our VEP extracts highly distinctive global and local feature representations from the Multi-Head Self-Attention (MSA) and Multi-Layer Perceptron (MLP) modules of ViT, integrating them with dynamic prompts to enhance knowledge learning and transferring. Additionally, we introduce a diversity loss mechanism to reduce redundant information accumulation in the learned prompts and enhance their discrimination, aiming to mitigate catastrophic forgetting effectively. Extensive experiments on benchmark datasets, including ImageNet-R, CUB200, and CIFAR100, validate the contribution of each component in VEP and demonstrate that VEP achieves state-of-the-art performance on the task of Continuous Learning while maintaining low parameter overhead. The source code of our designed VEP is at https://github.com/SWU-CSMediaLab/VEP.

Most existing Visible-Infrared Person Re-Identification (VI-ReID) methods primarily rely on fixed datasets for training, which struggle to accommodate continuously evolving cross-domain data, thus significantly limiting the adaptation in real-world dynamic scenarios. The task of Lifelong Visible-Infrared Person Re-Identification (LVI-ReID) emerged and is required to overcome the challenge of the semantic gap caused by both cross-modality and cross-domain data. Drawing inspiration from complementary learning systems, we propose a prompt-based dynamic learning framework to address the challenges inherent in LVI-ReID. Specifically, we design a Prompt Pool (PP) module to encapsulate shared knowledge across tasks or domains. In addition, we propose an instance-level prompt generator (IPG) to further enhance the model's ability to capture domain-specific knowledge, overcoming the limitations of a fixed-size prompt pool. For task-agnostic inference during the LVI-ReID phase, we develop a query-key mechanism that adaptively selects the most relevant prompt by evaluating the similarity between query tokens and keys, thereby addressing the nuanced requirements of varying tasks. Extensive experimental evaluations demonstrate the superiority of our proposed prompt learning-based PP-IPG framework over state-of-the-art methods in both lifelong learnings, lifelong person re-identification (LReID), and LVI-ReID settings. These results underscore the efficacy and practicality of our framework for advancing LVI-ReID across dynamic cross-modality and cross-domains. The source code of our designed PP-IPG method is at https://github.com/SWU-CSMediaLab/PP-IPG.

Visible-infrared person re-identification has been extensively explored, but it typically relies on stationary datasets for training. However, data is collected in a streaming manner in practical scenarios, necessitating the model's ability to continually learn without forgetting previous tasks. Existing methods focus simply on the lifelong single-modality person re-identification, but the visible images are sometimes unavailable, e.g., at night. To this end, this paper introduces a more challenging yet practical problem: Lifelong Visible-Infrared Person Re-identification (LVI-ReID). Inspired by the complementary learning systems, we propose a Tri-Token transformer with a Query-Key mechanism (TTQK) to tackle the LVI-ReID. Firstly, a general token is designed to capture robust domain-general features, shared across different domains, aiming to enhance the generalization capability. Subsequently, recognizing that different domains possess unique features like illumination and scenes, we allocate a specific token for each domain to extract significant domain-specific features, aiming to enhance the adaptability across domains. Furthermore, to prevent using the task identifier in the inference stage of LVI-ReID, we design a query-key mechanism to adaptively select the appropriate specific token based on the similarity between the query token and keys. Extensive experiments demonstrate that our method outperforms other lifelong learning and LReID methods. The source code of our designed LVI-ReID method is at https://github.com/SWU-CS-MediaLab/TTQK.

Few-Shot Class-Incremental Learning (FSCIL) aims to keep recognizing novel classes from a limited number of samples after training on abundant data from base classes while maintaining the performance of the old classes. The challenge, however, is that limited data from new classes not only leads to the issue of overfitting but also catastrophic forgetting. To address these two issues, we propose a causal inference strategy in the mainstream FSCIL framework, which encourages the model to learn significant knowledge in the base training session and enhance the model's ability to extract features to cope with the emergence of unseen classes in the incremental session, by improving the learning of causal relationships between features and predictions for perturbed samples. In addition, to improve the effectiveness of learning new tasks in the incremental sessions while preventing the model from overfitting to the novel class data, we freeze the feature extractor while adding a Fourier transform after the feature extractor in the incremental session. It can denoise the features, strengthen the features of the novel classes, and suppress the error in extracting the features of the limited number of samples directly from the feature extractor. Extensive experiments on CIFAR100, Caltech-USCD Birds-200-2011, and miniImageNet datasets show that our proposed framework achieves state-of-the-art performance on FSCIL. The source code of our designed framework is at https://github.com/SWU-CS-MediaLab/CIFSCIL.

Crowdsourcing is a useful and economic approach to data annotation. To obtain annotation of high quality, various aggregation approaches have been developed, which take into account different factors that impact the quality of aggregated answers. However, existing methods generally focus on single-label (multi-class and binary) tasks, and they ignore the inter-correlation between labels, and thus may have compromised quality. In this paper, we introduce a Multi-Label answer aggregation approach based on Joint Matrix Factorization (ML-JMF). ML-JMF selectively and jointly factorizes the sample-label association matrices collected from different annotators into products of individual and shared low-rank matrices. As such, it takes advantage of the robustness of low-rank matrix approximation to noise, and reduces the impact of unreliable annotators by assigning small (zero) weights to their annotation matrices. In addition, it takes advantage of the correlation among labels by leveraging the shared low-rank matrix, and of the similarity between annotators using the individual low-rank matrices to guide the factorization. ML-JMF pursues the low-rank matrices via a unified objective function, and introduces an iterative technique to optimize it. ML-JMF finally uses the optimized low-rank matrices and weights to infer the ground-truth labels. Our experimental results on multi-label datasets show that ML-JMF outperforms competitive methods in inferring ground truth labels. Our approach can identify unreliable annotators, and is robust against their misleading answers through the assignment of small (zero) weights to their annotation.