GroundWork is a labeling tool for satellite imagery. Anyone can create a custom machine learning training dataset using the Free or Pro version of the tool. Or, you can work with us to manage the labeling or modeling processes entirely.
For those with one discrete image to label, GroundWork’s current organizing unit, the “Project” is sufficient. However, we found it limiting for clients and colleagues who wanted to manage multiple images for their training dataset. Working with several areas of interest or grouping related images (say, of the same place at different times) was cumbersome. Take for example the training dataset we created for Azavea engineer James McClain’s cloud detection model. Comprised of 32 different images, this one case required the user to navigate four different pages:
With GroundWork’s new Campaign feature, managing our current research dataset is much easier:
Not only can I see how complete my set of projects is without opening them, but I can also now see all my related projects and their progress on one page. Of course, the project view still remains so you can check in on individual images as well. GroundWork users will be able to access the Campaign feature starting March 30. If you’re using the Free version, you’ll be able to add up to 10 Campaigns and upload 10 GBs of data.
In addition to Campaigns, our engineers have been working hard at enhancing GroundWork’s usability. Some recent features we’ve added are:
We're empowering organizations to
do more good.
- “Magic” wand: this tool detects and selects similar pixels in an image. Annotators working on segmentation projects can use it to create polygon labels without having to draw the outline of a feature.
- Flag: annotators can use this feature to identify tasks unsuitable for labeling (e.g., the task is cloudy), and project managers can then remove these tasks from the dataset or return it to the dataset for labeling if they believe the task was flagged in error.
- Split: some tasks contain too many features for one labeler to complete without fatiguing themselves. In such cases, the labeler can use this feature to split the task into four more manageable tasks. Split tasks can also be split further.
- Clickable task map: GroundWork users can now choose a task to label or validate by clicking on it in the task map. This is fantastic for times when you want annotators to focus on a specific portion of the task map!
Are you using GroundWork to create your machine learning training dataset? We’d love to hear from you about what you’re doing and what features might improve your experience.