Read some of our takeaways from training a data labeling team or tips for optimizing your labeling project on our company blog.
Our team helps you design the right type of annotation campaign, image processing, and tools for your use-case. Most importantly, our labeling lead works directly with you and the team at CloudFactory to make sure the project stays on track and that the labels are of the highest quality. Throughout the process, you’ll have GroundWork Pro access, which gives you more room to grow your projects.
It is then either accepted, modified, or rejected and sent back to the queue for relabeling. Validators are able to leave notes documenting any issues and giving instructions on needed adjustments. In our experience, validated labels can provide a significant boost to a machine learning model’s performance.
With a team dedicated to GroundWork clients, we can get started on your campaign within days as opposed to weeks without a long-term commitment.
The team of data analysts assigned to Azavea in July 2019. Photo by CloudFactory.
Our data analysts are seasoned GroundWork users with in-depth knowledge of its features, maximizing the efficiency of your labeling time. Many have worked with the tool from its inception and have played an integral role in suggesting and developing features that increase productivity and annotation quality. We communicate and problem-solve with them daily.
We worked with one of the largest international development organizations to create a dataset of urban tree canopy cover in Freetown, Sierra Leone.
In addition to our labeling services, we provided training, tracking, and feedback for their teams of managers and student labelers in Sierra Leone and Tanzania. The labeled data was used to train a feature detection algorithm.
We worked with this client to annotate several areas of interest with varying classes. This imagery, which is not on the traditional visual spectrum, provided several challenges.
In order to support their data, we provided custom visualization and added a multimodal image labeling feature to GroundWork. Our Labeling Lead and team of data analysts developed custom workflows and techniques to accommodate and improve the annotation of non-optical imagery.
Cloud cover is one of the most pernicious problems of satellite imagery. One of our machine learning engineers decided to tackle this problem by building a cloud detection model. Our professional annotaters produced over 20,000 labels for the model and learned to identify and annotate several different types of cloud cover. We also ran a cloud labeling competition for amateur labelers as part of a partnership with Radiant Earth. Within a week, participants from across the world had mapped around 2.3 million square kilometers of satellite imagery to create a signature training dataset and raise awareness of the SpatioTemporal Asset Catalog (STAC) spec.
With our customers, we’ve produced hundreds of thousands of labels on low- and high-resolution optical imagery as well as SAR imagery. Those labels include artificial and natural features, such as: cooling towers, solar panels, vehicles, storage tanks, and forests, bodies of water, clouds, and land cover.
Have labels that need to be cleaned up? Need help creating your machine learning models? We can do that, too!
Every labeling project begins with good imagery. In our latest white paper, we discuss cost, resolution, update frequency, third party vendors, and more. Read our free white paper to discover the best satellite imagery for your project.
Get the white paper
