The day my AI was right and I was wrong
You’re driving, your car’s lane assist is quietly keeping you centred, and then, without a sound, it stops helping. No warning. It just can’t see the lines anymore.
The paint has faded. And here’s the thing that surprised me: nobody really knows which streets are like that. Councils find out by sending special survey cars down every road, which is slow and expensive.
So for the OpenCV AI Competition 2026, I asked a simple question: could a computer figure it out from free street photos?
The first version lied to me
My first attempt looked brilliant. On the suburbs I built it with, it was almost always right. Then I tried it on suburbs it had never seen, and it did no better than flipping a coin.
That was a hard week. But it taught me the most important lesson of the whole project: if you only test on what you practised on, you’re just testing your own memory.
So I rebuilt it around a careful little inspector:
- It checks every photo first and throws out the blurry, dark or useless ones.
- It shows the good ones to an AI that can read pictures, and asks one question: can you see the lane lines?
- If it doesn’t have enough photos, it goes and finds more.
- If the answers disagree, it doesn’t guess. It hands the street to a person.
The plot twist
Get anna’s stories in your inbox
Join Medium for free to get updates from this writer.
To check its answers, I looked at every street myself, without seeing what the system said. On one test it seemed to make 8 mistakes. So I went back to those photos, ready to fix my code.
The lines were there. The system was right. I’d marked those photos rushed and tired, at 2 am, in a tiny window. I labelled every street again carefully, reported both rounds honestly, and built myself a better marking tool.
Sometimes the bug is the person checking the work.
Where it landed
On a final blind test of 88 places it had never seen, it agreed with me 72 times out of 75, and it passed 13 tricky ones to a human instead of guessing. It runs live on AWS, and running it cost me $0.
Try it yourself. Take a photo of any road from the driver’s seat and upload it. It checks it live in a few seconds:
d1dzzj0gvxxl4o.cloudfront.net
If you enjoyed this, I’d really appreciate your vote in the community round:
devpost.com/software/adas-ready-roads
I started this project wanting to teach a computer to see. I ended it having learned a lot more about how I see.
#OpenCVComp26 #ComputerVision #AWS #AI
