Train the Map AI
With just a few taps, you can teach AI to make maps.
Last updated
With just a few taps, you can teach AI to make maps.
Last updated
A map is a collection of symbols and images that stand for features of the real world, whether they are handwritten in chalk, printed in a book or digitized in an app. Those features change all the time. Roads are built. Speed limit signs are changed. Even the natural landscape changes; floods and landslides regularly change how people navigate the world.
Dashcams are the eyes that allow the Hivemapper Network to sense those changes. But as with human cognition, eyes are not enough. You need a brain. Map AI is the brain of the Hivemapper Network, extracting high-quality map data from images to help build fresher maps.
Hivemapper's Map AI algorithms automatically detect objects, such as speed limit signs, stop signs and traffic signals, and then position those objects on the map.
Hivemapper's community of human reviewers, known as "AI Trainers," complete bite-sized AI Trainer tasks to audit the outputs of these algorithms to ensure the data are accurate.
The results of these tasks are used for reinforcement learning from human feedback, refining the algorithms to do a better and better job of identifying objects and placing them in the world.
Check out this blog series to learn more about Map AI:
People who help improve Map AI are known as AI Trainers. Anybody with an internet connection can get started without permission or formal training. We typically see a few types of contributors:
Microtaskers: people who perform tasks for Amazon Mechanical Turk or Remotasks may find similarities in the work required for AI Training. These people often play the trainer games accurately and proficiently for long stretches of time in order to maximize their rewards.
Dashcam contributors: people who primarily contribute and earn rewards by collecting images often squeeze AI Training into free moments throughout their day to help the network succeed.
Casual gamers and map enthusiasts: people who enjoy games such as GeoGuessr and people who just love maps often find Hivemapper's trainer games to be a fun way to build a new type of map. It doesn't hurt they also earn HONEY rewards.
Visit Hivemapper AI Trainers from any desktop or mobile device.
Click the "Connect Wallet" button; this is required to earn rewards. If you are on a mobile device, tap the menu icon first and then the Connect Wallet button.
Sign in using Phantom or Breeze.
Go to the AI Trainers page and choose a task.
Auditing map data and training the Map AI is a complicated job, but Hivemapper has broken this work into simple, bite-sized tasks to allow anyone to participate. Each task focuses on a category of object that is important for mapmaking, such as speed signs, stop signs or roadwork zones.
The types of tasks will evolve over time based on the needs of the Map AI and map data customers. Over time, Map AI will be capable of handling more and more tasks entirely on its own, allowing AI Trainers to focus on tasks where human input is still needed.
There are a few different types of tasks that AI Trainers can do to teach the Map AI.
| Type of Game | Description | Objective | Skill Level | Examples |
|---|---|---|---|---|
Identify | Look at a mosaic and select all of the objects that meet the criteria | Teach the Map AI to correctly detect objects | Beginner | Stop Sign, Traffic Light, No Turns 1 |
Classify | Select the correct attribute for an object from a list | Teach the Map AI to correctly assign attributes | Beginner to Intermediate | Speed Limit, No Turns 2, Highway Exit Info |
Merge | Confirm whether two or more objects are the same exact object | Teach the Map AI to correctly merge objects seen more than once | Beginner to Intermediate | Merge |
Image QA | Determine whether images meet quality criteria | Teach the Map AI to detect poor mounts and image quality | Intermediate | Dashcam Tilt |
Position | Confirm or update the positioning and/or orientation of an object | Teach the Map AI to position and orient objects in physical space | Advanced | Position, Azimuth |
Some tasks require more skill than others. To ensure that skilled tasks are completed by skilled reviewers, the Hivemapper Network assigns a reputation score to each AI Trainer and sets a reputation threshold for each type of task.
Reputation is updated as Trainers complete tasks over time.
A good reputation is hard to build and easy to lose.
To access the most challenging and rewarding games, you need to be a 4-star AI Trainer or above.
Hivemapper has no tolerance for Trainers acting in bad faith and will aggressively protect the network against them.
AI Trainers share a weekly pool of HONEY rewards for their contributions. The rewards pool is divided among AI Trainers based on how many tasks they completed of each type, weighted by a reputation score that reflects the quality of their work.
AI Trainer reward amounts are subject to change and rewards are not guaranteed for Trainers who do not maintain a high reputation.
Each type of task has a weighting that is determined based on how long it takes to complete and the skill level required. More complex and more time consuming tasks receive higher rewards, incentivizing the most skilled AI Trainers to focus on tasks that have the greatest value for the map.
Reward weights are shown for each game on the AI Trainer dashboard, and are subject to change. Rewards are processed based on the weighting points the game had when you submitted reviews for the game.
We have a vibrant, active community of AI Trainers helping each other with tips, tricks and troubleshooting advice on Discord. Check out the #ai-trainers channel.
For questions which can't be answered by the community, please contact us at hi@hivemapper.com and our support team will be happy to assist you.
