10k crowdsourced images for training segnets
Learn more from the blog post, or on the comma.ai discord in the #comma-pencil channel.
It's 10,000 PNGs of real driving captured from the comma fleet. It's MIT license, no academic only restrictions or anything.
Run
./viewer.pyto see them with segnet overlay.
imgs/ -- The PNG image files masks/ -- PNG segmentation masks (update these!) imgs2/ -- New PNG image files paired with fisheye PNGs masks2/ -- PNG segmentation masks (update these!) segs/ -- The outputs in probability from our internal segnet (unreleased, too big)
1 - #402020 - road (all parts, anywhere nobody would look at you funny for driving) 2 - #ff0000 - lane markings (don't include non lane markings like turn arrows and crosswalks) 3 - #808060 - undrivable 4 - #00ff66 - movable (vehicles and people/animals) 5 - #cc00ff - my car (and anything inside it, including wires, mounts, etc. No reflections)
Start labelling! Useful label tools:
Fork this repository to your account using the "Fork" button in the top right
Create a new branch from the master branch, and use your labelling tool of choice to label some images
Open a pull request from your new branch to the master branch in the official repository to submit your changes!
Visit the #comma-pencil channel on the comma.ai Discord for the latest news and chat about the project.
comma10k is still a work in progress. For now, just cite the GitHub link. Once we reach 10k images, we'll release a paper, a train/test split, and a benchmark model.
For now, we are validating on images ending with "9.png" and are seeing a categorical cross entropy loss of 0.051. Can you beat this?
And it has been beaten with a CCE loss of 0.045, "comma10k-baseline" by YassineYousfi!
Can you beat that?