← Selected work
ExperimentsExperiment

Car Damage Detection Exploration

I explored YOLOv8 for detecting visible vehicle-body damage. The work available here focuses on converting COCO annotations for YOLO training and documenting the planned training and inference steps.

Overview

I explored YOLOv8 for detecting visible vehicle-body damage. The work available here focuses on converting COCO annotations for YOLO training and documenting the planned training and inference steps.

Objective

The experiment asks whether bounding boxes can identify visible damage regions in vehicle images for later review. Estimating repair cost or damage severity is outside its scope.

What I Built

  • A COCO-to-YOLO annotation-conversion workflow.
  • A proposed train/validation folder structure.
  • Dataset configuration for a damage-detection class.
  • Notebook material for data preparation.
  • Documentation covering training, inference, and result visualization.

Workflow

COCO images and annotations
  -> annotation conversion
  -> YOLO dataset structure
  -> train/validation configuration
  -> YOLOv8 training
  -> image inference and bounding boxes

Current Status

This remains an early learning experiment. The available files cover preparation and the intended workflow; they do not establish a finalized model, reproducible evaluation metrics, or deployment.

Limitations

  • Detection does not estimate repair cost, severity, or causality.
  • Dataset licensing and image-publication rights must be verified.
  • Performance depends heavily on viewpoint, lighting, damage type, and label consistency.
  • No production-readiness claim should be made without a separate held-out evaluation.

Technology Stack

Python, YOLOv8, Ultralytics, COCO annotations, Jupyter.

Publication Note

Before publishing examples, verify dataset licensing and image rights. If those rights are unclear, use newly created or explicitly licensed demonstration images instead.

Checks and current status

Not Yet Evaluated

Model assessment

Data-preparation files are available; a final model result is not established.

The work covers annotation conversion and the planned training steps. Evaluating a detector would require a held-out dataset, mAP and recall measurements, error analysis, and verified image rights.

Data preparation

AvailableCOCO-to-YOLO conversion and intended training workflow are documented.

Held-out evaluation

Not availableNo finalized reproducible metrics were found.

Deployment

Not applicableThe work remains an experiment.

Limitations: Detection does not estimate repair cost, damage severity, or causality.