PVIS-10K: A 10K drone image dataset with component-level annotations for rooftop photovoltaic inspection

PVIS-10K is a dataset of 10,065 undistorted RGB drone images with component-level annotations for rooftop photovoltaic (PV) inspection research. It covers 22 rooftop PV stations in Ningbo, Shaoxing and Shenzhen, China, and provides 3,516,675 image-level polygon annotations representing 63,632 distinct physical PV components. Multi-angle observations capture variations in installation height, projected component scale and orientation, dense arrangements, occlusion, shadows and reflections. Annotations were produced through three-dimensional component delineation, two-dimensional reprojection, and image-by-image inspection and correction. Each instance represents a single physical PV component, rather than an internal cell or an entire array. Partially occluded components retain their complete projected outlines, including occluded portions; fully occluded components are excluded. Annotation counts represent accumulated observations across images, not unique physical components. The dataset is partitioned by station: 12 Ningbo stations form the training set, while six Ningbo stations, three Shaoxing stations and one Shenzhen station form the test set. Both training and test images and annotations are provided. Independently downloadable station ZIP archives contain JPEG images and COCO-format polygon annotations, preserving the train/test directory structure. Supporting files include usage documentation, station-level split definitions and checksums. PVIS-10K supports component localization, delineation, image-based counting, and evaluation across unseen stations and cities. The dataset is distributed under the Creative Commons Attribution 4.0 International licence (CC BY 4.0). Data-processing and benchmark code is available at https://github.com/plzhang1998/PVDataset.