Siyuan Liu

Multibeam sonar is an important sensing modality for submerged reef detection, with potential applications in ROV operations, navigation safety, and offshore engineering. However, dedicated public acoustic datasets with object-level annotations for submerged reef or reef-like target detection remain limited, restricting the development and evaluation of relevant algorithms. To address this gap, we present SRAD, a multibeam acoustic dataset for submerged reef detection in a laboratory-recreated seafloor environment. The dataset contains more than 3,650 8-bit grayscale sonar images with corresponding ground-truth bounding-box annotations, together with native Oculus sonar logs and supporting data-processing tools. Data were collected in a controlled basin within a large-scale ocean engineering laboratory, where seafloor-related materials, including sand, gravel, natural rock reefs, algal reef materials, shell reefs, and simulated coral structures, were physically arranged to recreate a laboratory seafloor environment. To improve data consistency, an Oculus M750d multibeam sonar mounted on an ROV was used under consistent acquisition settings, and the ground-truth annotations were subjected to expert review and additional consistency checks. The resulting dataset provides labeled multibeam sonar images for the development, validation, and testing of submerged reef detection methods under controlled laboratory conditions, while the native Oculus sonar records and supporting processing tools enable further acoustic-data analysis and reuse.

Accurate near-field underwater visual perception plays a crucial role in marine ecological monitoring and benthic resource exploration. However, the superimposition of the biomimetic characteristics of benthic organisms and the optical degradation caused by the water medium frequently induces significant underwater visual camouflage phenomena. This causes the target foreground and the background to become highly fused in terms of color, texture, and structure, thereby severely limiting the performance of underwater vision algorithms in core tasks such as object detection, image segmentation, and 3D reconstruction. Existing public datasets predominantly focus on salient targets in clear water or under simple backgrounds, lacking comprehensive multi-task benchmarks explicitly tailored for visual camouflage scenarios. To address this gap, we constructed an Underwater Camouflaged Object Dataset (UCOD) for near-field benthic organisms, designed for multi-task analysis to jointly support image enhancement, object detection, pixel-level segmentation, and 3D scene reconstruction. The dataset comprises 7,000 high-resolution RGB images, including 3,500 images with detection annotations, 3,500 images with segmentation masks, and 16 reconstruction sequence folders for 3D reconstruction. It covers six categories of benthic organisms exhibiting typical camouflage characteristics: scallops, fish, conches, abalones, starfish, and sea cucumbers. During data acquisition, calibrated underwater imaging equipment was employed, and the kinematic parameters of the acquisition platform were controlled to improve the stability and spatial consistency of the collected data. The dataset provides useful training data and an evaluation basis for underwater multi-task perception research under visual camouflage conditions, while offering complementary data support for studies on underwater optical sensing and intelligent exploration.

Underwater object detection is crucial for marine ecological monitoring and resource exploration. However, low underwater contrast causes severe aliasing between foreground and background in the spatial domain, and conventional methods struggle to effectively decouple their features. Transformer architectures incur high computational overhead. Conversely, over-compressing these models severely degrades their ability to detect heavily camouflaged or small marine organisms. High-quality underwater samples are limited, and existing time-consuming generative strategies are hard to implement efficiently on edge devices. To address these challenges, this paper proposes a resource-efficient framework for underwater object detection. First, we design a Wavelet-Enhanced Feature Pyramid Network that combines a saliency-focus mechanism and a discrete wavelet transform to overcome background noise in both spatial and frequency domains, extracting features of hidden small objects. Second, a data-dependent dynamic token pruning technique removes redundant tokens, effectively mitigating the computational bottleneck without sacrificing essential semantic capacity. Finally, for extreme sample scarcity, we introduce a Feature Correction Module and a two-stage fine-tuning and feature correction strategy, using a high-precision teacher model to guide a compressed student network in adaptively compensating for optical shifts with few samples. Experiments on URPC2020 and DUO demonstrate that our method improves small object detection accuracy while reducing parameter count and computational overhead, striking a good balance between accuracy and inference efficiency.

With the development of acoustic and optical exploration technologies, devices such as optical camera, laser scanner and multibeam sonar provide the possibility of underwater accurate perception. The optimization of acoustic and optical data through algorithms have become one of the major concerns in underwater computer vision field. However, these optimization and improvement algorithms require a large amount of underwater multimodal data for training and evaluation. To address these needs, we propose a multimodal acousto-optic dataset for underwater image enhancement, detection, and reconstruction (MAOUD), which was collected from our underwater simulation environment and includes optical RGB images, acoustic images, labeled images, laser point clouds and acoustic videos. To ensure the accuracy of the dataset, we used a state-of-the-art underwater multimodal integrated detector with guaranteed corresponding kinematic parameters. The dataset can be used for training and evaluation work for a variety of underwater acoustic and optical tasks, serving as a standardized training and validation benchmark for multimodal underwater acoustic and optical algorithms, which holds significant importance for advancing underwater exploration technology.

The 3D reconstruction of vessel hulls is crucial for enhancing safety, efficiency, and knowledge in the maritime industry. Neural Radiance Fields (NeRFs) are an alternative to 3D reconstruction and rendering from multi-view images; particularly, tensor-based methods have proven effective in improving efficiency. However, existing tensor-based methods typically suffer from a lack of spatial coherence, resulting in gaps in the reconstruction of fine-grained geometric structures. This paper proposes a spatial multi-scale weighted NeRF (MDW-NeRF) for accurate and efficient surface reconstruction of vessel hulls. The proposed method develops a novel multi-scale feature decomposition mechanism that models 3D space by leveraging multi-resolution features, facilitating the integration of high-resolution details with low-resolution regional information. We designed separate color and density weighting, using a coarse-to-fine strategy, for density and a weighted matrix for color to decouple feature vectors from appearance attributes. To boost the efficiency of 3D reconstruction and rendering, we implement a hybrid sampling point strategy for volume rendering, selecting sample points based on volumetric density. Extensive experiments on the SVH dataset confirm MDW-NeRF’s superiority: quantitatively, it outperforms TensoRF by 1.5 dB in PSNR and 6.1% in CD, and shrinks the model size by 9%, with comparable training times; qualitatively, it resolves tensor-based methods’ inherent spatial incoherence and fine-grained gaps, enabling accurate restoration of hull cavities and realistic surface texture rendering. These results validate our method’s effectiveness in achieving excellent rendering quality, high reconstruction accuracy, and timeliness.

Reconstructing medical images from partial measurements is a critical inverse problem in computed tomography (CT) and magnetic resonance imaging (MRI). While diffusion models excel in 2D medical imaging, their extension to 3D reconstruction remains challenging due to high computational costs. In this work, we achieve efficient 3D reconstruction from sparse views by reusing 2D diffusion models that are pre-trained unconditionally on 2D slices (CT for the CT tasks, MRI for the MRI tasks), without any 3D or task-specific training. Our method integrates curvature-regularized model-based iterative reconstruction with the denoising steps of diffusion models via the Alternating Direction Method of Multipliers (ADMM), where a z-axis curvature prior enforces inter-slice consistency and a variable-sharing scheme keeps the per-iteration cost close to that of purely 2D sampling. Experimental results show improved fine-detail and structural fidelity under sparse-view and limited-angle CT, with up to 15% higher PSNR and SSIM than baselines, and competitive MRI reconstruction across diverse datasets, including out-of-distribution settings. The proposed framework thus offers a computationally efficient route to high-fidelity 3D reconstruction, with promising clinical applications such as low-dose CT and accelerated MRI.