Open AI models that assist road inspection
Weights for detection models that run on a phone, published under a license that allows commercial use (Apache-2.0). Usage notes and evaluation results are included, so you can use them as they are or retrain them on your own data.
Technology that eases inspection gets better faster the more people use it. So the models that detect road damage, signs, curve mirrors and vehicles are published rather than kept in-house.
Published models
-
Curve mirrors
Curve Mirror YOLOX-Tiny
Finds curve mirrors in images taken while driving.
Available
-
Road signs
Road Sign YOLOX-Tiny
Finds road signs in images taken while driving and sorts them into five classes.
Available
-
No-snow-throwing signs
No-Snow-Throwing Sign YOLOX-Tiny
Finds where snowplows must stop and may resume throwing snow.
Available
-
Vehicles
Vehicle RF-DETR Small
Finds vehicles in video, sorts them into nine types and counts them by road.
Available
-
Potholes
Pothole RF-DETR Nano
Finds potholes from a phone mounted in a car.
Available
Curve mirror photo: Kamikou / Mapillary (CC BY-SA 4.0)
Road sign photo: pika_sky / Mapillary (CC BY-SA 4.0)
Intersection footage: SHIBUYA COMMUNITY NEWS (sibch.tv) / Internet Archive (CC BY 3.0)
License
Free to use, including commercially
Weights and code are released under Apache-2.0. You can build them into your work or products, and modify and redistribute them. Usage and required notices are in the README and NOTICE files where they are published.
What has started, and what is next
- 01
Model releases
Published on Hugging Face.
View on Hugging Face - 02
Leaderboard
Submitted models will be scored under the same conditions on private evaluation images, and ranked on a leaderboard.
- 03
Commissioned app development
Apps built around the open models, developed on commission.
See AI App Development
FAQ
Q.01Where can I download them?
Weights and documentation (README) can be downloaded together from the infra-intelligence page on Hugging Face. Each model's page also links to its release.
Q.02Is there a cost?
No. Downloading and using the weights is free.
Q.03Can you build an app using the open models?
Yes. Phone and web apps built around the open models are developed on commission.
Q.04Can I build them into my work or products?
Yes, under the terms of Apache-2.0. Note that evaluation sets are small and real-road performance has not been measured yet. Test on images from where you will use them first.
Q.05Can I retrain them on my own data?
Yes. The PyTorch weights load directly in the rfdetr library, so they make a good starting point for fine-tuning.
Q.06Will other damage or road furniture be covered?
Expanding to cracks and other damage, and to more road furniture such as signs, is under consideration. Any news will be posted on this page.
Start by describing your situation
Whether it is a MICHIKARTE trial or an AI app, your plans do not need to be final. Describe the photos or video you have and what you are struggling with, and you will get an email reply on what can and cannot be done. The contact form is in Japanese, but you are welcome to write in English.