Underwater image restoration is challenging due to light absorption and scattering, which cause color distortion, low contrast, and uneven illumination. Existing methods either ignore the underwater imaging process or rely on simplified assumptions, especially treating background light as a spatially uniform component, which limits restoration performance in complex environments. Recent diffusion-based approaches improve restoration quality by exploiting strong image priors, but their physical modeling of background illumination remains insufficient. In this paper, we propose a physics-constrained diffusion method with spatially varying background light for underwater image restoration. The proposed method explicitly models background light by decomposing it into illumination intensity and mixed chromaticity components, and incorporates physical consistency constraints into the diffusion sampling process. A staged optimization strategy is further introduced to progressively optimize physical parameters while keeping the diffusion prior fixed during inference. Experiments on multiple underwater datasets demonstrate consistent improvements in both full-reference and no-reference metrics. The current implementation requires approximately 568 s per 256×256 image and is therefore more suitable for high-quality offline restoration than real-time applications.
Visual emotion recognition aims to infer human emotional states from visual cues. Adapting pretrained Vision–Language Models (VLMs) to this task remains challenging. In-Context Learning (ICL) enables a frozen model to condition its predictions on a few labeled demonstrations without task-specific fine-tuning. Here, demonstrations are retrieved from an existing labeled support pool; the small prompt size does not imply a small total annotation budget. Despite this potential, how to effectively leverage ICL for visual emotion recognition remains underexplored. In this paper, we propose EmoICL, a hybrid retrieval framework that leverages VLMs to enhance emotion recognition. To this end, we first introduce Emotion-Aware Dual Retrieval (EDR), which generates emotion-sensitive descriptions using guided prompts and refines retrieved samples based on visual similarity and cross-modal relevance. Then, we design Category-Aware Diversity Optimization (CDO), which combines category-aware filtering with Maximum Marginal Relevance to reduce label-level interference and improve visual diversity. Experiments on three benchmark datasets show that EmoICL achieves the best average performance across different VLM backbones, with a mean absolute gain of approximately 1.96 percentage points over the strongest baseline by average accuracy for each backbone. Code is available at: https://github.com/Atlantic4Li/EmoICL.
Most existing GAN inversion methods aim to strike a good trade-off between fidelity and editability. However, a pre-trained GAN latent space only encodes information from in-domain regions while the input image often involves out-of-domain regions. As a result, extending the original latent space by encoding out-of-domain regions to improve fidelity will negatively affect the editability of the model. To address this, we propose a novel dual dynamic hypergraph-driven fusion (DHFusion) method, which consists of a dual dynamic hypergraph-CAM network (DH-Net) and an editing-driven fusion network (EF-Net). Specifically, DH-Net first employs the differential activations between the initial inverted and edited images to dynamically construct dual hypergraphs from the perspectives of short-range and long-range spatial dependencies. In this way, high-order relationships between attribute-relevant regions are effectively modeled, enabling our model to generate an accurate and comprehensive edit-aware mask for locating the edited regions. Subsequently, EF-Net leverages this mask as weights to perform editing-driven feature-level fusion between the original image and the initial edited image, generating high-fidelity edited images with the reduced ghosting effect. Extensive quantitative and qualitative experiments demonstrate that our method outperforms several state-of-the-art methods. Our work clearly shows the potential of dual dynamic hypergraphs for GAN inversion. Our code is available at https://github.com/Huang-KT/DHFusion.
The pursuit of large receptive fields has shaped the evolution of computer vision frameworks, spanning from convolutional neural networks (CNNs) to Transformers and Mamba. Recently, large-kernel operations have revitalized CNNs, making them competitive once again and widely applicable across diverse vision tasks. However, scaling kernel sizes inevitably results in substantial growth in both parameters and computational overhead. Consequently, existing approaches are often limited to small kernels or resort to decoupled designs for large kernels. In this paper, we propose a simple and efficient large kernel network for image restoration, termed ArtIR, motivated by the channel redundancy observed in image restoration models. Specifically, ArtIR applies adaptive large-kernel operations to a collapsed single channel and employs an ultra-lightweight channel attention mechanism to restore channel diversity. To complement local features, we further introduce a large kernel fusion module that integrates multi-scale information. Unlike most prior methods that focus on a narrow set of restoration tasks, we comprehensively evaluate ArtIR across single-degradation, all-in-one, and composite degradation scenarios. Beyond generic restoration, we also assess our model on domain-specific applications such as ultra-high-definition restoration, medical imaging, and remote sensing. Extensive experiments demonstrate that ArtIR achieves state-of-the-art performance while maintaining high efficiency and fast inference.
3D semantic occupancy prediction (OCC) is a crucial task in vision-centric autonomous driving, focusing on reconstructing fine-grained 3D geometry and semantics of the scene surrounding the ego vehicle. Despite recent progress, OCC still faces a significant challenge in generalizing to real-world scenarios due to the unpredictable domain shifts between training data and diverse testing environments. A promising solution to this challenge is test-time adaptation (TTA), which enhances model generalization by updating its parameters to align with the testing domain during inference. However, applying existing TTA methods to OCC is non-trivial as: i) Practical deployment of OCC on resource-constrained onboard devices imposes requirements on memory efficiency. ii) OCC must rapidly adapt to dynamic scenes in latency-sensitive autonomous driving scenarios, posing a challenge for fast test-time adaptation. To address the above issues, we propose a novel fast and memory-efficient Occupancy Test-Time Adaptation (OccTTA) method. Specifically, i) to achieve memory-efficient TTA, we propose a novel TTA framework for OCC based on lite-boost units; ii) to enable fast TTA, we propose a novel forward-backward adaptation strategy that synergizes computationally lightweight forward updates with selective backward optimization. Specifically, our strategy selectively triggers backward updates only for divergent inputs, while employing a k-NN classifier to perform efficient forward updates for accelerating the model’s responsiveness to domain shifts, thereby simultaneously improving adaptation efficiency and achieving superior performance. Moreover, to address the class imbalance problem in OCC, we introduce an adaptive class-balanced loss that dynamically prioritizes challenging classes, enabling more robust TTA. Extensive experiments on the nuScenes-C benchmark verify the effectiveness of our method.
