What is comma10k?

comma10k is a crowdsourced dataset of 10,000 real driving images designed to help train AI systems that understand what a car's camera sees on the road. The project is run by comma.ai, a company working on autonomous driving technology. The idea is that by collecting and carefully labeling real driving footage, researchers and developers can build better computer vision models that distinguish between roads, lane markings, other vehicles, pedestrians, and areas where a car simply shouldn't drive. The dataset works by pairing raw driving photos with corresponding "segmentation masks", essentially overlay images where every pixel is color-coded by category. The project defines five categories: road, lane markings, undrivable areas, movable objects like cars and people, and the driver's own vehicle. A car equipped with a dashcam captures the images, and then human volunteers manually trace over each image to identify what every section of the picture represents. There's a viewer script included that lets you see the images with the AI-generated overlays applied, giving you a sense of what a finished segmentation model produces. The people who would use this are machine learning researchers, self-driving startups, or hobbyists building computer vision systems for automotive applications. For example, a small team building a lane-keeping assistant could train their model on this data to recognize road boundaries, or a student working on a robotics project could use it to teach a system how to spot pedestrians and other vehicles in real-time driving scenarios. The data is released under a permissive MIT license with no academic-only restrictions, so commercial projects can use it freely. What makes this project interesting is its crowdsourced approach to data labeling, which is typically one of the most expensive and time-consuming parts of building AI systems. Volunteers claim images through a shared spreadsheet, label them using browser-based tools or image editors like GIMP and Photoshop, and submit their work through pull requests. The community coordinates through Discord, and there's even a video tutorial to help newcomers get started with the labeling process.