Folder Structure

After downloading and unzipping, you'll find a data/ folder with:

• train/ - training images by label.
• test/ - evaluation set for final metrics.
• val/ - validation set for tuning.

Labels

Each label is a subfolder name representing a category. Keep labels snake_case for portability (e.g., cracked_asphalt).

Image Format

All images are .jpg using tuned compression for high visual quality while minimizing size.

Train / Val / Test Split

Default is 60 / 30 / 10. Use the sliders on the download page to choose any ratio that sums to 100. Each image is assigned randomly at pack time, so re-downloading produces a fresh split.

Image Augmentation

Pick any combination of augmentations (rotate, flip, crop, blur, noise, color jitter, and more) and use the percentage slider to control what fraction of train images get randomly augmented — from 0% up to 100%. Augmented files are suffixed _aug.

Loading in PyTorch

Point torchvision.datasets.ImageFolder at the unzipped data/train and data/val directories. Pair with transforms.Normalize using the ImageNet mean / std printed in the bundled meta.json if you're fine-tuning a pretrained model.

Custom B2B Datasets

Need a production-grade dataset with custom classes, QA, and YOLO / COCO annotations? Our team builds licensed datasets on demand — see For Business for scope and timeline.