TrashCan

TrashCan 数据集由带注释的图像（目前为 7,212 张图像）组成，其中包含对垃圾、ROV 和各种海底动植物的观察。此数据集中的注释采用实例分割注释的格式：包含掩码的位图，标记图像中的哪些像素包含每个对象。 TrashCan 中的图像来自日本海洋地球科学技术局 (JAMSTEC) 策划的 J-EDI（JAMSTEC 深海图像电子图书馆）数据集。该数据集包含来自 JAMSTEC 自 1982 年以来主要在日本海中运营的 ROV 的视频。该数据集有两个版本，TrashCan-Material 和 TrashCan-Instance，分别对应不同的对象类配置。最终目标是开发适用于车载机器人部署的高效准确的垃圾检测方法。虽然之前已经创建了包含海洋环境中垃圾边界框级别注释的数据集，但据我们所知，TrashCan 是第一个实例分割注释的水下垃圾数据集。我们希望该数据集的发布将有助于进一步研究这一具有挑战性的问题，使海洋机器人社区更接近于解决自动垃圾检测和清除这一紧迫问题的解决方案。

The TrashCan dataset comprises 7,212 annotated images capturing observations of marine debris, Remotely Operated Underwater Vehicles (ROVs), and various benthic marine flora and fauna. Annotations within this dataset follow the instance segmentation annotation format: bitmap masks that identify which pixels in an image correspond to each individual object. The images used for the TrashCan dataset are sourced from the J-EDI (JAMSTEC Deep-Sea Image Electronic Library) dataset, curated by the Japan Agency for Marine-Earth Science and Technology (JAMSTEC). This J-EDI dataset contains video footage collected by ROVs operated by JAMSTEC primarily in the Sea of Japan since 1982. There are two variants of the TrashCan dataset: TrashCan-Material and TrashCan-Instance, which correspond to distinct object class configurations. The overarching goal of this dataset is to support the development of efficient and accurate debris detection methods suitable for deployment on on-board robotic systems. While prior datasets have been created with bounding-box-level annotations for marine debris in underwater environments, to the best of our knowledge, TrashCan is the first underwater debris dataset with instance segmentation annotations. We anticipate that the release of this dataset will facilitate further research on this challenging problem, bringing the marine robotics community closer to developing solutions for the pressing issue of automated marine debris detection and removal.

搜集汇总

背景与挑战

背景概述

TrashCan是一个水下垃圾检测图像数据集，包含7,212张带实例分割注释的图像，这些图像源自日本海洋地球科学技术局的深海图像库，覆盖垃圾、ROV和海底动植物。该数据集有两个版本（TrashCan-Material和TrashCan-Instance），旨在支持开发高效准确的垃圾检测方法，是首个提供实例分割注释的水下垃圾数据集，适用于海洋机器人社区的自动检测研究。