A robot-view image dataset of photovoltaic panel reference lines with paired synthetic soiling
Description
This dataset contains 6,200 robot-view photovoltaic panel images for reference-line perception under clean, synthetic soiling, and laboratory real-soiling conditions. Each 160×120 RGB image is paired with line-instance annotations for white cell-gap lines and horizontal/vertical frame–cell border lines. Synthetic soiling spans five levels from very light to extreme, so every clean source image appears under six conditions in total. The dataset is designed for line instance detection, heading angle estimation, and reversal triggering in real-time navigation of photovoltaic cleaning robots. It includes paired clean-to-soiled synthetic images with unchanged line geometry, plus 200 laboratory real-soiling images collected on separate panels. Reproducible code is provided for synthetic soiling generation. The dataset is built to evaluate how well structural line perception holds up at the input size used for onboard real-time inference when the same panel-line geometry is exposed to increasing soiling severity, alongside a real-contamination subset to test physical generalization. The paired synthetic subsets show how soiling affects line visibility, while the real-soiling set tests whether models trained on synthetic soiling generalize to physical dirt. Annotations are formatted as line-instance labels rather than dense semantic masks: ρ and θ define the line in Hough space, endpoint coordinates bound the visible segment, and class labels denote structural roles. The dataset can thus be used to compare model robustness across controlled soiling levels and to study real-time visual navigation of photovoltaic cleaning robots. Note: This version replaces the near-duplicate scenes, leaving no duplicate pair across the training, validation, and testing splits. We also double cheack all 14,956 line annotations, dropping stray short segments and keeping same-class lines parallel within each image.