Synthetic Interior Damage Dataset (SIDD)

The Synthetic Interior Damage Dataset (SIDD) dataset consists of 663 annotated images captured inside the cabin of a Citroën Ami using a single fisheye camera mounted on the vehicle ceiling, providing a 180° field of view that covers the entire interior. Synthetic damages were digitally added to the seat surfaces, including the seat base and backrest areas, simulating realistic wear, scratches, and tears.

The dataset is designed to support research in computer vision, anomaly detection, and damage localization tasks within automotive interiors. Each image is annotated in YOLO detection format and prepared for use in supervised machine learning experiments. The dataset contains 478 labeled instances, and all annotations correspond to a single class: 0 – damage.

The dataset also includes a helper script, split_yolo_dataset.py, which can be used to divide the dataset into training, validation, and test sets. The script:

Copies all images to the corresponding split folders.

Copies existing YOLO-format label files to the same splits.

Creates empty label files for images without annotations, allowing them to serve as negative examples.

Ensures the split proportions match the configured ratios (default: 70% train, 20% validation, 10% test).

This script facilitates dataset organization for training YOLO object detection models and ensures that images without synthetic damages are properly handled.