MARS: Mask Attention Refinement for Car Damage Segmentation

Welcome to the official repository for MARS—an innovative deep learning model for precise car damage instance segmentation. By leveraging advanced self-attention mechanisms with sequential quadtree nodes, MARS significantly improves segmentation accuracy compared to methods like Mask R-CNN, PointRend, and Mask Transfiner.

At MARSAIL, we built the AI with a clear mission in mind: to revolutionize the car insurance and automotive repair industries by automating the estimation of labor costs and spare parts. By leveraging cutting-edge AI technology, MARS can analyze vehicle damage with remarkable precision, providing insurance companies and garages with faster, more accurate assessments. Our goal is to empower the auto industry to lead in efficiency, innovation, and accuracy, setting a new standard for automation that drives cost savings, enhances customer satisfaction, and ultimately transforms how insurance and repair services operate.

Key Achievements:

+1.3 maskAP improvement with the R50-FPN backbone
+2.3 maskAP improvement with the R101-FPN backbone on the Thai car-damage dataset

MARS was showcased at ICIAP 2023 in Udine, Italy.

Why MARS is the Best

Here’s why MARS stands out in the field of car damage segmentation:

Our advanced model significantly outperforms traditional methods in detecting and segmenting car damage. The following figure demonstrates the superior segmentation results achieved by MARS:

The image above showcases the precise damage segmentation, validating why MARS outperforms other models such as Mask R-CNN, PointRend, and Mask Transfiner.

📚 MARSAIL 2025 Research Drop: Shaping the Future of Automotive AI

We are excited to unveil MARSAIL’s latest research breakthroughs in automotive AI for 2025. Our team has been pushing the boundaries of AI to bring precision, automation, and efficiency to the car insurance and repair industries. Here’s a glimpse of our newest research:

🌟 Vision Research (2025)

ALBERT Advanced Localization and Bidirectional Encoder Representations from Transformers for Automotive Damage Evaluation Read on arXiv A transformer-based model for more accurate and faster car damage assessments, improving the insurance claims process.
SLICK Selective Localization and Instance Calibration for Knowledge-Enhanced Car Damage Segmentation in Automotive Insurance Read on arXiv A precision-driven approach to car damage segmentation, enhancing the performance of AI models in automotive insurance and repair.

🧠 NLP Research

DOTA Deformable Optimized Transformer Architecture for End-to-End Text Recognition with Retrieval-Augmented Generation Read on arXiv Revolutionizing document processing and OCR recognition for automotive services, ensuring more efficient insurance claim handling.

🌟 Throwback Highlight (2023)

MARS Mask Attention Refinement with Sequential Quadtree Nodes for Car Damage Instance Segmentation Read on Springer The foundational model that set the stage for all modern car damage segmentation solutions.

At MARSAIL, our mission is simple: To lead the way in AI-powered innovation for the automotive insurance and repair industries. These papers represent just the beginning of our journey to automate damage assessment, accelerate claims processing, and redefine the future of vehicle repair. Stay tuned for more groundbreaking advancements!

Publications

If you're interested in the academic work behind MARS, please check out the following publication:

Quick Start

Installation Instructions

Follow these steps to set up the MARS model:

git clone https://github.com/kaopanboonyuen/MARS.git
cd MARS
python3 -m venv mars-env
source mars-env/bin/activate
pip install -r requirements.txt

Visit GitHub Repository

Datasets

Our models were trained on both public and private datasets:

Public Dataset
Private Dataset: Restricted access due to licensing agreements with THAIVIVAT INSURANCE PCL.

Citation

If you find our work helpful, please consider citing it:

@inproceedings{panboonyuen2023mars,
  title={MARS: Mask Attention Refinement with Sequential Quadtree Nodes for Car Damage Instance Segmentation},
  author={Panboonyuen, Teerapong, et al.},
  booktitle={International Conference on Image Analysis and Processing},
  year={2023},
  organization={Springer}
}

Dataset Citation

If you're utilizing the public dataset Car Damage Detection (CarDD), which includes 4,000 high-resolution images and over 9,000 well-annotated instances across six damage categories (dent, scratch, crack, glass shatter, lamp broken, and tire flat), please make sure to cite the following paper:

@article{wang2023cardd,
  title={Cardd: A new dataset for vision-based car damage detection},
  author={Wang, Xinkuang and Li, Wenjing and Wu, Zhongcheng},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  volume={24},
  number={7},
  pages={7202--7214},
  year={2023},
  publisher={IEEE}
}

Live Demos

Experience MARS in action: Visit Live Demo