Currently, the assessment of vehicle damage involves evaluating the extent of damage and estimating the initial repair costs. This process is often slow because it requires multiple steps of examination examination steps by expert personnel, using numerous photos of the damaged vehicles. This results in long processing times, potential errors, and biases in the evaluation, To address these issues, a web application has been developed to predict vehicle damage. However, the training datasets for these models are limited and hard to obtain, as vehicle image datasets often have inappropriate angles, poor lighting conditions, low resolution, or noise such as excessive reflections, making it difficult to identify vehicle damage accurately.
To solve these problems, a system using Blender to create realistic vehicle image datasets from 3D vehicle models has been developed. This dataset is used to train the model, which is then integrated into a web application capable of predicting the cost and extent of damage. The AI employs deep learning to understand different levels of damage and provides output in the form of damage assessment and repair cost estimation.
The main benefit of developing this web application for damage assessment is that it can reduce delays and biases in evaluating vehicle accident damage. It can assess which parts are damaged, the extent of the damage and the damage’s extent, and estimate the repair costs, Users can share the assessment data and results with insurance companies or relevant parties, enabling faster processing.
For example, users can provide photos of the damaged parts, damage levels, and estimated costs to insurance companies, allowing for quicker compensation processing. Additionally, the assessment history and other details can also be recorded for future reference or analysis.
Our project has received funding support and has successfully advanced to the final round of the National Software Contest (NSC) Thailand 2024.
Additionally, we have been honored with the Thailand New Gen Inventors Award 2025 (I-NEW GEN AWARD 2025) in the **Higher Education Level - Tourism and Creative Economy** category, receiving a **Silver Medal**.
Our project, DentAnalyzer: A Web Application for Predicting Car Accident Damage Costs, integrates synthetic image generation and machine learning for damage assessment. Below are key research studies that provide insights into our approach.
This study explores the use of synthetic data for AI training, allowing for a diverse, accurate, and large-scale dataset without human dependency. It introduces a procedural technique to simulate car damage on 3D models and generate labeled 2D images. Read more about CrashCar101
TEXTure is a model designed for applying and transferring textures from images to 3D models based on textual descriptions, enhancing the realism of generated surfaces. Read more about TEXTure
This research focuses on automatically generating viewpoint-labeled datasets, aiding object detection and orientation estimation models. Read more about Unsupervised
ShapeNet is a large-scale repository of 3D models with category labels, widely used for AI training in 3D shape recognition. Visit ShapeNet
This dataset contains high-quality car images with fine-grained labels, enabling AI models to distinguish between similar vehicle types. Read more about Large-Scale Car Dataset
CarDD provides a dataset specifically designed for training deep learning models to detect and classify car damage. Read more about CarDD
This study presents an AI model trained to classify different types of car damage based on image input, enhancing the accuracy of automated damage assessment. Read more about Car Damage Detection and Classification
@inproceedings{itccscc2025dentanalyzer,
author = {Phuris Kruacharee and Thanaphorn Kanking},
title = {DentAnalyzer: Web Application for Predicting Car Damage from a Single Image and Generating Realistic Car Images for AI Training},
booktitle = {2025 International Technical Conference on Circuits/Systems, Computers, and Communications (ITC-CSCC)},
address = {Seoul, Korea, Republic of},
publisher = {IEEE},
year = {2025},
month = {Jul.},
doi = {10.1109/ITC-CSCC66376.2025.11137611},
url = {https://ieeexplore.ieee.org/document/11137611}
}