第八届全球校园人工智能算法精英大赛-无人机低空航拍图像语义分割专题赛训练数据集

## 2026第八届全球校园人工智能算法精英大赛-无人机低空航拍图像语义分割专题赛-数据集（示例数据）
随着低空经济的蓬勃发展，无人机技术在测绘、安防、农业监测及城市规划等领域的应用日益广泛。无人机低空航拍图像具有视角独特、分辨率高、地物细节丰富等特点，但同时也面临着视角变化大、目标尺度差异显著、背景复杂等挑战。语义分割作为计算机视觉的核心任务之一，旨在对图像中的每个像素进行分类，是实现无人机视觉感知智能化的关键。
然而，现有的通用语义分割模型在直接应用于低空航拍场景时，往往因训练数据与航拍数据分布差异（Domain Gap）而表现不佳。此外，航拍图像中不同地物类别（如农田、建筑、道路）的分布极不均衡，且测试场景中某些类别的占比可能与训练场景存在显著偏差（如特定区域农田占比过高），这要求模型具备极强的泛化能力和鲁棒性。本赛题希望通过开发高效的无人机低空航拍图像语义分割模型，探索在复杂分布数据环境下提升模型性能的方法，推动无人机视觉技术的落地应用。
数据集文件元信息以及数据文件，请浏览“数据集文件”页面获取。

### 数据集及数据说明
数据来源于公开无人机数据集（包含部分遥感图像）及南京航空航天大学工信部重点实验室自收集的无人机低空航拍数据。所有数据经过统一清洗、类别映射及预处理。
#### 图像规格：
所有图像已统一切分为 1024*1024 像素的 patch，并进行了重新排序。
#### 类别体系：
共包含 9 个语义类别（含背景及忽略类），具体包括：农田、水体、荒地、车辆、忽略、背景、建筑、道路、森林。每张图片会同时出现多个语义类别。
#### 数据分布：
训练集共 6996 张图像。测试集分为三次发布，测试集 1含500 张图像，将被用于为初赛测试数据；测试集 2含1300张图像，将被用于为复赛测试数据；测试集 3含1794 张图像，将与测试集1、2合并，一起用于为半决赛测试数据。
#### 分布特性：
数据集存在显著的类别分布不均现象。统计显示，部分类别（如农田、水体）在测试集中的占比可能远高于训练集或全局平均水平，这对模型的类别平衡性提出了较高要求。例如，某些场景下超过一半的测试图片可能集中在特定地物类别上。
#### 数据格式：
训练数据将按照图像文件夹和标签文件夹的形式给出，标签图为单通道灰度图，像素值对应类别 ID。

### 下载方法
:modelscope-code[]{type="sdk"}
:modelscope-code[]{type="git"}

2026 8th Global Campus AI Algorithm Elite Competition - UAV Low-Altitude Aerial Image Semantic Segmentation Track (Sample Data)

With the booming development of low-altitude economy, UAV technology has been increasingly applied in fields such as surveying and mapping, security, agricultural monitoring, and urban planning. UAV low-altitude aerial images feature unique perspectives, high resolution, and rich ground object details, but also face challenges such as large perspective variations, significant differences in target scales, and complex backgrounds. As one of the core tasks of computer vision, semantic segmentation aims to classify each pixel in an image, which is the key to realizing intelligent UAV visual perception.

However, when existing general semantic segmentation models are directly applied to low-altitude aerial scenarios, they often perform poorly due to the distribution difference (Domain Gap) between training data and aerial data. In addition, the distribution of different ground object categories (such as farmland, buildings, roads) in aerial images is extremely uneven, and the proportion of certain categories in the test scene may deviate significantly from that in the training scene (e.g., an excessively high proportion of farmland in a specific area), which requires the model to have strong generalization ability and robustness. This competition aims to develop efficient UAV low-altitude aerial image semantic segmentation models, explore methods to improve model performance in complex distributed data environments, and promote the practical application of UAV vision technology.

For the metadata and data files of the dataset, please browse the "Dataset Files" page.

### Dataset and Data Description
The data is sourced from public UAV datasets (including some remote sensing images) and UAV low-altitude aerial data collected by the Key Laboratory of Ministry of Industry and Information Technology, Nanjing University of Aeronautics and Astronautics. All data has undergone unified cleaning, category mapping, and preprocessing.

#### Image Specifications:
All images have been uniformly cropped into 1024*1024 pixel patches and reordered.

#### Category System:
A total of 9 semantic categories (including background and ignore class) are included, specifically: farmland, water body, wasteland, vehicle, ignore, background, building, road, forest. Multiple semantic categories will appear in each image.

#### Data Distribution:
The training set consists of 6996 images in total. The test set is released in three batches: Test Set 1 contains 500 images, which will be used as the preliminary contest test data; Test Set 2 contains 1300 images, which will be used as the semi-final contest test data; Test Set 3 contains 1794 images, which will be merged with Test Set 1 and 2 for the final contest test data.

#### Distribution Characteristics:
The dataset exhibits significant class imbalance. Statistics show that the proportion of some categories (such as farmland and water bodies) in the test set may be much higher than that in the training set or the global average, which imposes high requirements on the class balance of the model. For example, in certain scenarios, more than half of the test images may be concentrated on specific ground object categories.

#### Data Format:
The training data will be provided in the form of image folders and label folders. The label maps are single-channel grayscale images, where pixel values correspond to category IDs.

### Download Method
:modelscope-code[]{type="sdk"}
:modelscope-code[]{type="git"}

搜集汇总

背景与挑战

背景概述

该数据集为第八届全球校园人工智能算法精英大赛的无人机低空航拍图像语义分割专题赛训练数据，包含统一处理为1024×1024像素的航拍图像，涵盖农田、水体、建筑等9个语义类别。数据集具有显著的类别不平衡特性，训练集含6996张图像，测试集分三个阶段共3594张图像，旨在推动针对无人机低空图像的语义分割模型发展，提升模型在复杂数据分布下的性能。