Telenav open-sourced the machine learning based sign detection platform that powers the automatic detection of nearly 100 sign types in the OpenStreetCam images you contributed. You can already see these detections in the latest version of the OpenStreetCam JOSM plugin to help you map, and iD integration will come soon as well.
Machine learning gets better with training. The more known instances of a particular sign that are fed into the system, the more reliable the automatic detections for that sign type will become.
Our Map Team has spent thousands of hours manually tagging and validating traffic signs in images, and the resulting training data is open source as well. But did you know you can help improve the detection system yourself as well? Let us show you how.
If you go to the trip details on the OpenStreetCam web site, you will see three ‘tabs’ on the left. The first one takes you to the main trip info. The second one takes you to an OSM edit mode, that lets you quickly go over detections and see if they need to be added to OSM. (Separate post! The third tab is the sign validation mode. If the tab icon has a number with it, there are unverified signs to work on.
The bottom part of the screen shows all detected signs. The ones that have been validated already will have a green checkmark with them. The ones that have been invalidated will have a red ‘X’.
You can validate or invalidate the automatic detection if the sign on the image exactly matches / doesn’t match the automatic detection, by clicking the corresponding button on the left.
Power Validator Workflow
You can validate entire trips with many detected signs very quickly by using some of the power functions available:
- Next to the trip slider, underneath the image, you will find a small magnifying glass button. Clicking this will automatically zoom and pan the image to the detection
- Use Cmd (Mac) / Alt (Windows / Linux) and the left and right arrows to quickly jump to the next detection
- Use Cmd / Alt up and down to validate or invalidate the currently highlighted detection.