A Multimodal Optical Dataset for Underwater Image Enhancement, Detection, Segmentation, and Reconstruction.

Area of Science:

• Computer Vision
• Robotics
• Marine Biology

Background:

• Precise underwater environmental perception relies on multimodal devices, integrating optical cameras and LiDAR.
• Underwater computer vision requires robust algorithms for enhancing RGB images and laser point clouds.
• Extensive, well-organized, and labeled underwater multivariate data are essential for training and evaluating these algorithms.

Purpose of the Study:

• To introduce the Multimodal Optical Underwater Dataset (MOUD), a comprehensive resource for underwater AI tasks.
• To facilitate advancements in underwater image detection, segmentation, enhancement, and 3D reconstruction.
• To support the development of more accurate underwater sensing and exploration technologies.

Main Methods:

• Collected over 18,000 original RGB images and 12,000 labeled images from underwater simulation scenarios.
• Acquired 60 point cloud sets and several labeled point clouds, featuring nine distinct marine object classes.
• Utilized a state-of-the-art image-laser underwater detector with precise kinematic parameters for data validation.

Main Results:

• The MOUD dataset contains a substantial volume of diverse underwater visual and 3D data.
• Labeled images include annotations for objects such as scallop, starfish, conch, holothuria, seaweed, coral, reef, abalone, and barnacle.
• The dataset is validated for accuracy and utility, suitable for training and evaluating advanced underwater perception algorithms.

Conclusions:

• The MOUD dataset significantly addresses the need for high-quality multivariate data in underwater computer vision.
• It provides a crucial resource for advancing research in underwater image enhancement, detection, segmentation, and 3D reconstruction.
• This dataset is pivotal for enhancing the capabilities of underwater optical exploration and sensing technologies.