As a leader in high-quality training data, Sama supports clients across various use cases and applications. The ability to identify specific keypoint landmarks and track their relationship to one another is unlocking some of the most interesting developments in computer vision technology. This includes motion tracking like human pose identification for virtual fitness trainers or sports analytics. It also includes facial landmarks for emotion analysis, facial verification security, or driver alertness detection. It could also include hand gesture control for AR/VR or sign language transcription.
We are thrilled to announce our support for custom keypoint shapes in our training data platform trusted by the world’s leading AI teams, for vector image and video annotation. While Sama has long supported other vector shapes like bounding boxes, polygons, and cuboids, these keypoint use cases require custom shapes that have a predefined number and order of points. Sama can now be configured to support skeletons, hands, eyelids or any other complex custom shape. Each keypoint can have its own label, color, and connection to other points. Multiple keypoint shapes are supported on the same annotation project.
Optimizing for Quality and Efficiency
This new capability optimizes quality and efficiency in producing ground truth training data for our clients. Our expert annotators have a facilitated drawing experience where the shape builds itself as they annotate each point. The correct label for each point is inferred from the order that it is placed so no extra time is spent assigning labels to each point. We see quality improvement from visualizing the annotations as a cohesive connected shape instead of as free standing individual points. The custom keypoint shape also always has the same number and order of points—this means that no points could ever be omitted or mislabeled. In addition to the quality benefits, our A/B testing showed an 27% decrease in annotation time on a skeleton image annotation workflow over individual points.
Completed annotation data is returned to our clients in a structured delivery format that contains information about each point, its relative order in the array and its connection to other points.
If you’re currently working on a computer vision algorithm that requires high quality keypoint annotations on images or videos, make sure to read more on our platform page or get in touch with an expert to discuss your training data needs.
