HPRT-YOLO dataset Part3-val

HPRT-YOLO Dataset (A fused UAV maritime search-and-rescue object detection dataset) This dataset is designed for small-object detection in UAV-based maritime search and rescue (SAR) and water-surface surveillance scenarios. It is formed by integrating two public datasets: SeaDronesSee and An Ensemble Deep Learning Method with Optimised Weights for Drone-Based Water Rescue and Surveillance. • Source Dataset 1: SeaDronesSee It contains 14,211 maritime UAV images with 6 original labels: ignored, swimmer, boat, jetski, life-saving appliance, buoy, and provides an official train/val/test split of 8,914/1,547/3,750. The data covers diverse coastal/offshore/harbor environments and includes challenges such as complex sea states, glare, and background clutter. • Source Dataset 2: Drone-Based Water Rescue and Surveillance It includes 3,613 images (train/val/test: 2,766/333/514) annotated with 6 classes: human, wind/sup-board, boat, kayak, buoy, sailboat. To build HPRT-YOLO, we performed unified data cleaning, label alignment, and class remapping across the two sources, and obtained a fused dataset with 8 classes: ignored, human, boat, jetski, buoy, life-saving appliance, wind/sup-board, kayak. To improve generalization under complex maritime conditions, we apply random cropping, scaling, flipping, rotation, and Mosaic augmentation during training; augmentation is disabled for validation and inference. Because the dataset is relatively large, it is uploaded here in separate parts. The following is Part 3, which contains the val set. Note: This is a curated and fused derivative of public datasets. Users should cite the original datasets and their corresponding papers/pages and comply with the original licenses.HPRT-YOLO 数据集（一种融合式无人机海上搜救目标检测数据集） 本数据集面向基于无人机的海上搜救（Search and Rescue, SAR）及水面监视场景下的小目标检测任务设计，通过整合两个公开数据集SeaDronesSee与《An Ensemble Deep Learning Method with Optimised Weights for Drone-Based Water Rescue and Surveillance》构建而成。 • 源数据集1：SeaDronesSee 该数据集包含14211幅海上无人机图像，涵盖6类原始标注：忽略区域（ignored）、游泳者（swimmer）、船只（boat）、喷气滑雪艇（jetski）、救生设备（life-saving appliance）、浮标（buoy），并提供官方预设的训练/验证/测试集划分：8914/1547/3750。数据覆盖多样的近岸、远海及港口环境，包含复杂海况、眩光、背景杂波等典型挑战场景。 • 源数据集2：Drone-Based Water Rescue and Surveillance 该数据集包含3613幅图像，训练/验证/测试集划分比例为2766/333/514，标注类别共6类：人类（human）、冲浪板/桨板（wind/sup-board）、船只（boat）、皮划艇（kayak）、浮标（buoy）、帆船（sailboat）。 为构建HPRT-YOLO数据集，我们对两个源数据集开展了统一的数据清洗、标签对齐与类别重映射操作，最终得到包含8类目标的融合数据集：忽略区域（ignored）、人类（human）、船只（boat）、喷气滑雪艇（jetski）、浮标（buoy）、救生设备（life-saving appliance）、冲浪板/桨板（wind/sup-board）、皮划艇（kayak）。 为提升模型在复杂海事条件下的泛化能力，我们在训练阶段采用随机裁剪、缩放、翻转、旋转及马赛克（Mosaic）数据增强策略；验证与推理阶段则禁用数据增强操作。 由于本数据集体量较大，我们将其分卷上传，本文档为第三部分，仅包含验证集。 注意：本数据集为公开数据集的经筛选与融合的衍生版本，使用者需引用原始数据集及其对应论文/页面，并遵循原始开源许可协议。