MyUnderwaterWorld: Intelligent Underwater Scene Representation

人类对海洋的探索已经有数百年的历史，活动仍在继续，但它们总是受到潜水专家数量、技术，尤其是成本的限制。先进的成像技术可以将水下发现转移给岸上具有特定知识的专家，如地质学家、考古学家和生物学家。从这些图像序列中进行三维（3D）重建可以增强对水下生物、物体和海床的理解。然而，如果没有几个月的强化和处理时间，目前的解决方案还不能提供高分辨率和清晰度的水下3D表示。这主要是因为数据的限制和计算的复杂性，显然，由于失真、光的后向散射和浑浊条件，处理水下环境的序列是具有挑战性的。MyUnderwaterWorld项目旨在为基于人工智能（AI）的开发提供水下图像的深入分析，从而形成一个新的框架，用于图像质量增强和水下场景的高分辨率3D场景表示，其中包含海底和感兴趣的物体。我们假设3D场景可以使用定义良好的先验知识从原始水下数据中准确而直接地建模。这可以通过描述多样化和可靠的水下数据集来实现。我们将实时视觉SLAM和稀疏辐射场分层结合，并使用基于水下先验知识开发的新型损失函数进行训练。这将提高3D表示的质量，并提供更有效和灵活的工作流程。它还将为后续基于机器的处理提供更强大的特征提取，并为交付提供更有效的压缩。

The oceans have been explored for hundreds of years and the activities still continue, but they are always limited by the number of diving experts, technologies and in particular costs. Advanced imaging enables transferable underwater discovery to onshore experts with specific knowledge required, such as geologists, archaeologists and biologists. Three-dimensional (3D) reconstruction from these image sequences enhance understanding of underwater organisms, objects and the seabed. However, current solutions have not yet provided high resolution and definition of underwater 3D representation without months of intense enhancements and processing time. This is mainly because of limitation of data and computational complexity as, obviously, processing the sequences of underwater environments is challenging due to distortion, backscatter of light and turbidity conditions. MyUnderwaterWorld project aims to provide intensive analysis of underwater imagery for Artificial Intelligence (AI)-based development, leading to a novel framework for image quality enhancement and high-resolution 3D scene representation of underwater scene, which contains seabed and objects of interest. We hypothesise that the 3D scene could be modelled accurately and directly from raw underwater data using well-defined prior knowledge. This could be achieved by characterising diverse and reliable underwater datasets. We will combine real-time visual SLAM and sparse radiance fields hierarchically, trained with a novel loss function developed from prior knowledge of underwater. This will improve quality of 3D representation, and offer more efficient and flexible workflows. It will also facilitate more robust feature extraction for subsequent machine-based processing and more efficient compression for delivery.