geohot/img-labeler

img-labeler is a web-based tool that helps people label images for training self-driving car AI. Instead of installing software, you just open a link in your browser, paste in a picture URL, and start marking up the image to show the AI what's what, cars, lanes, roads, and your own vehicle. The workflow is hands-on and visual. You pick a color category (like "movable" for cars and motorcycles, or "lane markings"), then click around objects to trace their outlines. You can zoom in for precision, adjust layer opacity, and arrange layers, for instance, sending the road layer behind everything else. Anything you don't label gets automatically filled in as "undriveable" when you save, so you don't have to paint every last pixel. The tool also crops your work automatically and encourages overlapping layers to avoid gaps at the edges. This tool was built specifically for contributors to comma10k, an open-source project by comma that's collecting thousands of dashcam images annotated for self-driving research. A volunteer might grab a batch of images, trace masks around cars and lanes using this labeling tool, then upload their results back to the comma10k repository via a pull request. It's crowd-sourced data annotation, many people each labeling a few images to build a large, high-quality dataset together. What's notable here is the simplicity of the approach. Rather than requiring people to install specialized annotation software or set up a development environment, this project runs entirely in the browser. That lowers the barrier to contribution significantly, someone with basic computer skills and a GitHub account can participate in building self-driving AI training data. The tradeoff is that it's narrowly scoped: the color categories and workflow are tailored to comma10k's specific annotation needs, so it wouldn't work out-of-the-box for other types of image labeling projects.