Car Damage Dataset (CDD) Overview
- Car Damage Dataset (CDD) is a curated image dataset with precise annotations for various damage types and vehicle parts, supporting computer vision research.
- Methodologies across CDD variants include bounding boxes, polygon masks, and synthetic augmentation, enabling benchmark testing for detection and segmentation tasks.
- Applications of CDD span automated insurance claims, fraud prevention, and fine-grained damage segmentation, driving improvements in model robustness and domain adaptation.
A Car Damage Dataset (CDD) is a curated image dataset with ground-truth annotations for various types of visual damage, defects, and vehicle parts captured under real-world or simulated conditions. Such datasets are critical for developing, benchmarking, and validating computer vision models for automated vehicular damage assessment, insurance claim automation, fraud detection, and related tasks. Multiple datasets bearing the name or abbreviation "CDD" have been published since 2018, exhibiting significant variation in scope, annotation granularity, and benchmarking protocols. The following sections synthesize the defining characteristics, methodologies, and research applications of these datasets as reported in peer-reviewed preprints and dataset documentation.
1. Dataset Scope, Structure, and Variants
Car Damage Datasets have been released with variable scope, damage taxonomy, and annotation protocols, reflecting evolving goals in damage identification, fraud claims, and automated assessment.
- The first CDD, released in (Li et al., 2018), comprises ≈2 170 images: 1 790 web-scraped images of single damages and ≈380 images from 92 distinct vehicles in a parking-lot, each imaged multiple times for claim/fraud scenarios. Damage is categorized as scratch, dent, or crack with bounding-box (⟨xmin, ymin, xmax, ymax⟩) annotations, no segmentation masks.
- A more recent variant, named CDD in the context of (Panboonyuen, 12 Jun 2025), includes 12 000 high-resolution images annotated with 26 real damage types, 7 fake-damage types, and 61 vehicle part classes. This dataset uses instance-level polygon segmentation (COCO-style JSON) with mask and bounding box per object or region and is partnered with an extensive annotation protocol.
- Other published datasets (e.g., the "CDD" in (Baig et al., 21 Aug 2025)) focus on narrow domains such as minor dent detection, with 2 241 images, single-class bounding-box annotations, and highly controlled acquisition protocols.
- Reference datasets for comparison, such as CarDD (Wang et al., 2022) and CrashCar101 (Parslov et al., 2023), cover 4 000–101 050 images, broader class sets (up to 6 real types in CarDD, 5 in CrashCar101), and polygon-level segmentation (CarDD) or synthetic, pixel-accurate masks (CrashCar101).
| Dataset (citation) | Size | Damage Classes | Annotation Type | Modes/Protocols |
|---|---|---|---|---|
| CDD (Li et al., 2018) | ≈2 170 | 3 (scratch, dent, crack) | Bounding box | Real/fraud protocols |
| CDD (Panboonyuen, 12 Jun 2025) | 12 000 | 26 real, 7 fake, 61 parts | Polygon masks (COCO) | Multi-task, multi-label |
| CDD (Baig et al., 21 Aug 2025) | 2 241 | 1 (dent) | Bounding box (YOLOv8) | Single-class |
| CarDD (Wang et al., 2022) | 4 000 | 6 | Polygon (COCO/SOD) | Detection/segmentation |
| CrashCar101 (Parslov et al., 2023) | 101 050 | 5 | Pixel segmentation | Synthetic, domain random. |
2. Damage Taxonomies and Annotation Schemes
The definition of damage classes and annotation detail is a key differentiator among datasets.
- Class cardinality ranges from 1 (only "dent" in (Baig et al., 21 Aug 2025)) to 26 (comprehensive damage list in (Panboonyuen, 12 Jun 2025)). Major common classes include scratch, dent, crack, and various breakages.
- Fake-damage annotation appears in (Panboonyuen, 12 Jun 2025), distinguishing between real and intentionally simulated/painted defects, supporting anti-fraud model development.
- Vehicle part segmentation is included only in datasets such as (Panboonyuen, 12 Jun 2025) (61 classes) and synthetic CrashCar101 (Parslov et al., 2023) (27–61 classes), facilitating joint part/damage reasoning.
- Annotation format evolves from bounding boxes (Li et al., 2018, Baig et al., 21 Aug 2025) to polygon masks (Wang et al., 2022, Panboonyuen, 12 Jun 2025), enabling both detection and dense segmentation tasks. YOLO and COCO JSON schemas are prevalent.
This suggests that advances in annotation complexity (polygons, part-damage links, fake/real distinction) reflect rising ambitions for multi-task models and practical deployment robustness.
3. Data Acquisition, Environments, and Augmentation
Acquisition protocols and dataset diversity directly influence model generalization:
- Sources: Datasets use web scraping (Google/Bing/Baidu in (Li et al., 2018); Flickr/Shutterstock in (Wang et al., 2022)), in-situ claim documentation (Panboonyuen, 12 Jun 2025), and synthetic rendering (Parslov et al., 2023).
- Environments: Images span indoor studios, outdoor parking, roadside, natural and studio lighting, with controlled versus "in-the-wild" settings. (Li et al., 2018) simulates repeat-claim fraud with multiple parking-lot views per damage.
- Preprocessing: Common steps include EXIF-based auto-orientation, resizing to square frames (e.g., 640×640 or 1024×1024), JPEG/PNG formats, and normalization protocols (mean/std reported in (Panboonyuen, 12 Jun 2025)).
- Augmentation: Real-time augmentations (horizontal flip, brightness/contrast, mosaic, MixUp) are widely used (Baig et al., 21 Aug 2025, Wang et al., 2022). Synthetic datasets (Parslov et al., 2023) exploit domain randomization via environment maps, car colors, and procedural damage.
Dataset heterogeneity in environmental context and acquisition (real vs. synthetic) is a recognized driver of domain generalization for segmentation and detection models.
4. Benchmarking Protocols and Evaluation Metrics
Evaluation metrics and benchmarking tasks are adapted to detection, segmentation, and retrieval-focused use cases:
- Detection and Segmentation: Standard object detection and instance segmentation metrics are used: Intersection-over-Union (IoU), precision/recall (), F1-score, mean Average Precision (mAP), and mask AP (e.g., COCO-style, (Wang et al., 2022, Panboonyuen, 12 Jun 2025)).
- Fraud Detection/Retrieval: Older protocols (Li et al., 2018) apply nearest-neighbor matching with deep feature fusion (VGG-16, histograms) and rank-based retrieval accuracy (Rank-1, Rank-10).
- Small-object handling: Segmentation benchmarks distinguish AP_S (for small objects), showing significant gains from enhancements such as multi-scale augmentation and focal loss (Wang et al., 2022).
- Few-shot Sim2Real: Synthetic datasets (Parslov et al., 2023) demonstrate improved mIoU under few-shot fine-tuning, quantifying domain adaptation efficacy.
- Multi-task Learning: Accuracy, precision, recall, and F1-score per class/task are published for both damage and part segmentation (e.g., ALBERT-V9D/V9P in (Panboonyuen, 12 Jun 2025)).
| Task | Metric | Cited Example |
|---|---|---|
| Damage Detection | Precision, Recall, [email protected] | (Baig et al., 21 Aug 2025, Wang et al., 2022) |
| Instance Segmentation | AP_mask, AP_50, AP_S | (Wang et al., 2022, Panboonyuen, 12 Jun 2025) |
| Sim2Real Transfer (few-shot) | mIoU | (Parslov et al., 2023) |
| Fraudulent Claim Retrieval | Rank-1, Rank-10 Accuracy | (Li et al., 2018) |
| Part Segmentation | mIoU, Accuracy, F1 | (Parslov et al., 2023, Panboonyuen, 12 Jun 2025) |
5. Applications and Model Benchmarking
CDD variants support a range of computer vision tasks fundamental to automotive inspection:
- Insurance claim automation: Visual evidence-based verification and expedited assessment (Li et al., 2018, Panboonyuen, 12 Jun 2025).
- Fraud detection: Detection of repeated/fake claim submissions (Li et al., 2018, Panboonyuen, 12 Jun 2025).
- Fine-grained damage/part segmentation: Joint part-damage understanding supporting downstream cost estimation, repair triage, and fleet monitoring (Panboonyuen, 12 Jun 2025, Parslov et al., 2023).
- Model benchmarks: Deep architectures evaluated include YOLO variants (Li et al., 2018, Baig et al., 21 Aug 2025), Mask R-CNN, DCN⁺, SegFormer, and transformer-based ALBERT (Wang et al., 2022, Parslov et al., 2023, Panboonyuen, 12 Jun 2025).
- Data-driven robustness studies: Sim2real transfer, augmentation, and class-imbalance handling are explicit targets (Wang et al., 2022, Parslov et al., 2023, Baig et al., 21 Aug 2025).
Baseline results are available for each variant. For example, on (Panboonyuen, 12 Jun 2025), the ALBERT-V9D model achieves 0.9472 accuracy and 0.8926 F1 on 26-class damage classification; Mask R-CNN, DCN, and DCN⁺ yield AP_mask of 49.4, 52.5, and 57.0, respectively, on CarDD (Wang et al., 2022).
6. Data Access, Licensing, and Limitations
Data accessibility and use restrictions vary substantially:
- (Li et al., 2018): Access only via direct author contact; licensing unspecified but presumably limited to academic use.
- (Panboonyuen, 12 Jun 2025): Non-commercial academic/industry license; data available on request and agreement.
- (Baig et al., 21 Aug 2025): Published on Zenodo (https://doi.org/10.5281/zenodo.16654735); publicly available with citation requirement.
- (Wang et al., 2022): Fully public at https://cardd-ustc.github.io; COCO-style segmentation and detection annotations.
- (Parslov et al., 2023): Synthetic data can be generated or accessed per the procedural pipeline.
Identified limitations include constrained class coverage (single-class in some variants), absence of segmentation masks or stereo data, modest size relative to future needs, and persistent domain gaps between synthetic and real-world images. Recommendations across recent literature emphasize multi-class, multi-modality (mask, bounding box, keypoint), fake-damage realism, and improved environmental diversity for robust insurance, autonomous, and fleet applications.
7. Comparative Context and Future Directions
CDD and its analogues are part of an expanding corpus of car damage datasets that enable comprehensive benchmarking and new methodologies in vision-based automotive inspection.
Key trends:
- Explicit modeling of fake/tampered damage (Panboonyuen, 12 Jun 2025) for fraud resilience.
- Joint annotation of part, damage, and context, enabling holistic multi-task reasoning.
- Synthetic data (e.g., CrashCar101 (Parslov et al., 2023)) to overcome annotation bottlenecks and support sim2real transfer studies.
- Improved augmentation, small-object detection, and segmentation approaches, including focal loss, boundary-aware SOD modules, and domain-adaptive pretraining (Wang et al., 2022, Parslov et al., 2023).
- Consistent advocacy for expansion to include more comprehensive defect taxonomies, pixel-level localization, stereo/3D data, and extreme/rare-case representation (Baig et al., 21 Aug 2025, Wang et al., 2022).
Researchers are encouraged to select datasets and protocols aligned with task requirements—fraud detection, fine-grained segmentation, sim2real adaptation—and to reference the original dataset sources for detailed annotation schema, licensing, and access conditions (Li et al., 2018, Panboonyuen, 12 Jun 2025, Wang et al., 2022, Parslov et al., 2023, Baig et al., 21 Aug 2025).