Ruralscapes

我们收集了 20 部描绘农村地区的高质量 4K 视频。 Ruralscapes 包括各种景观、不同高度的不同飞行场景和跨越广泛尺度的物体。视频序列长度从 11 秒到 2 分 45 秒不等。该数据集由 17 分钟的无人机飞行组成，共有 50,835 个完全注释的帧，分为 12 个类别。其中有 1,047 个是手动注释的，每秒一次。据我们所知，它是最大的真实无人机视频语义分割数据集。
标签提供了很好的细节水平，但是由于小物体的空间分辨率降低，准确的分割很困难。由于数据集的性质，某些类别（例如 haystack）非常小，而其他类别（例如人）也具有特写镜头。手动标记是一个耗时的过程。根据来自 Liceul "Petru Cercel" din Targoviste 的 21 名志愿者的反馈，他们对数据集进行了分割，他们平均需要 45 分钟来标记整个帧。这意味着分割手动标记的 1,047 帧需要 846 小时。

We collected 20 high-quality 4K videos depicting rural areas. The Ruralscapes dataset encompasses diverse landscapes, varying flight scenarios at different altitudes, and objects spanning a wide range of scales. The video sequences range in length from 11 seconds to 2 minutes and 45 seconds. This dataset consists of 17 minutes of drone footage, with a total of 50,835 fully annotated frames categorized into 12 classes. Of these, 1,047 frames were manually annotated at one frame per second. To the best of our knowledge, this is the largest real-world drone video semantic segmentation dataset. The labels provide a high level of detail, but accurate segmentation is challenging due to the reduced spatial resolution of small objects. Due to the nature of the dataset, some categories (e.g., haystacks) are extremely small, while others (e.g., humans) also include close-up shots. Manual annotation is a time-consuming process. According to feedback from 21 volunteers from Liceul "Petru Cercel" din Targoviste who segmented the dataset, they required an average of 45 minutes to annotate a full set of frames. This means that annotating the 1,047 manually labeled frames would take 846 hours.

搜集汇总

背景与挑战

背景概述

Ruralscapes是一个专注于农村地区的无人机视频语义分割数据集，包含20部4K视频、50,835个注释帧和12个类别，是目前最大的真实无人机视频数据集。数据集的特点是具有多样化的景观和飞行场景，但标注过程耗时且某些类别样本较少。