FPVCrosswalk2025: A dataset for first-person view crosswalk segmentation in adverse weather and lighting conditions
Description
This dataset supports crosswalk segmentation for use in assistive navigation technologies for the visually impaired. It includes 3000 synthetic and 300 real-world first-person view (FPV) images, each paired with a binary segmentation mask. The synthetic set was generated using a fine-tuned Stable Diffusion model, with prompts covering general and varied environmental conditions (sunny, cloudy, rainy, and night). The real-world images are distributed across the same environmental conditions and extracted from video recordings of pedestrians approaching crosswalks. Each crosswalk appears in no more than two images from different directions to ensure diversity. All images were manually annotated using a custom tool that defines crosswalks as quadrilateral regions. The dataset is organized by source (synthetic/real) and environmental conditions, with matching folder structures for images and masks. It can be used for training and evaluating computer vision models, exploring synthetic data augmentation, and improving assistive systems for visually impaired pedestrians.
Files
Institutions
- Sveuciliste Josipa Jurja Strossmayera u Osijeku Elektrotehnicki Fakultet