🚘 CarDD Dataset

CarDD is a novel, public, large-scale dataset specifically designed for vision-based car damage detection and segmentation.

The dataset contains 4,000 high-resolution car damage images with over 9,000 well-annotated instances, making it the largest public dataset of its kind.

The high resolution of the images (average 684,231 pixels) is a key advantage over existing datasets that have a much lower average resolution (50,334 pixels). Higher resolution allows for more detailed annotations and the potential to detect finer damages.

CarDD Dataset Overview and Features

CarDD features six common external car damage categories, chosen based on frequency of occurrence and clear definitions from insurance claim statistics.

1. Dent

2. Scratch

3. Crack

4. Glass shatter

5. Tire flat

6. Lamp broken

Annotation process

The annotation process involved experts from the car insurance industry and trained annotators following specific guidelines based on insurance claim standards.

These guidelines address challenges like

• mixed damages (priority rules)

• damages across components (boundary splitting)

• adjacent same-class damages (boundary merging).

For object detection and instance segmentation, the annotations include masks and bounding boxes associated with each of the six damage types.

Each instance has a unique ID, category information, mask contours, and bounding box coordinates, following the COCO dataset format.

For SOD, pixel-level binary ground truth maps are provided.

Dataset splits

The dataset is split into training (70.4%), validation (20.25%), and test (9.35%) sets, maintaining a consistent ratio of instances for each category across the splits.

Near-duplicate images were explicitly removed to prevent data leakage.

Uses

The dataset provides comprehensive annotations for multiple computer vision tasks, including:

*   **Classification:** Identifying the type of damage.

*   **Object Detection:** Locating the damaged regions with bounding boxes.

*   **Instance Segmentation:** Precisely outlining the damaged areas with pixel-level masks.

*   **Salient Object Detection (SOD):** Identifying the damaged regions as salient objects through binary maps.

CarDD presents several challenges for model development due to the nature of car damage:

*   **Fine-grained distinctions** between damage types like dents and scratches.

*   **Diversity in object scales and shapes** of the damages.

*   A **significant proportion of small objects**, particularly for dent, scratch, and crack categories.

*   The fact that damages like **dent, scratch, and crack can be intertwined and visually similar**.

Availability

The CarDD dataset is publicly available at https://cardd-ustc.github.io.

However, access requires agreeing to the license terms of Flickr and Shutterstock, as the dataset does not own the copyright of the images.

The dataset is intended for non-commercial research and educational purposes.

Measures were taken to protect user privacy by mosaicking or deleting faces and license plates.

Citation

@ARTICLE{CarDD, author={Wang, Xinkuang and Li, Wenjing and Wu, Zhongcheng}, 
journal={IEEE Transactions on Intelligent Transportation Systems}, 
title={CarDD: A New Dataset for Vision-Based Car Damage Detection}, 
year={2023}, 
volume={24}, 
number={7}, 
pages={7202-7214}, 
doi={10.1109/TITS.2023.3258480}}