We are excited to announce new Alegion Platform feature releases: a major update to our image annotation capabilities including a new version of SmartPoly, a one-click segmentation tool.
Alegion has enhanced the user experience of Alegion Image annotation with new tooling for entity relationships, pose estimations, and skeletal keypoint connections. With the successful launch of these features in Alegion Video, we are now bringing these features to our image annotation platform to enable the creation of more sophisticated and accurate data sets.
Entity Relationships via Hierarchical Entities
Hierarchical entities, like identifying how objects are related to specific individuals in images, are now possible with entity relationships tools. Using pre-configured workflows, we enable you to identify, for example, which laptop and chair belongs to people sitting around a conference room or which person is shopping for a given set of groceries. A menu automatically pops up when a lower hierarchical element is labeled, enabling you to establish the object’s relationship to a parent entity.
Pose Estimation with Skeletal Keypoint Connections
Our new pose estimation feature for image annotation augments our existing keypoint segmentation tooling. Now, annotators can add keypoints and define the relationships between them. Lines will automatically connect keypoints to help annotators visualize the defined relationships and ensure annotation accuracy. Within both video and image annotation, pose estimation is a computer vision technique to predict human movements. Use cases might be detecting a medical issue when body parts are out of alignment, predicting who will win a sports match, or analyzing whether an athlete is training efficiently.
Enhanced JSON Result Format
Alegion Image annotation has upgraded its JSON result format to make it easier to inspect and use in scripting.
SmartPoly is a worker augmentation tool within the machine learning ecosystem in Alegion Platform, which also includes pre-labeling, work deflection, object tracking. SmartPoly speeds up the rate of annotation by locating the whole object with a single click. It is available in both video and image annotation, and works by harnessing a convolutional neural network (CNN) to segment the object and render a polygon. Points automatically appear on the polygon outline for manual refinement. Since the model is classless, it can be deployed without training in a wide variety of use cases.
Alegion Platform now uses an enhanced version of SmartPoly, a one-click segmentation tool to identify objects in images and video. Additional clicks can add or remove pixels from the auto-captured area.
In this enhanced version of SmartPoly, sliding bar controls allow users to refine the captured area if too many or too few pixels are captured. For area estimation control, an underlying machine learning model returns a value for each pixel in the clicked area and a confidence level that the pixel is or is not part of the object. Another sliding bar control allows the user to select how many points they want to appear on the object’s outline for greater refinement. During tests run with the original SmartPoly, there was a 30% reduction in annotation time, enabled by a 0.5 second polygon detection after click and a high accuracy rate resulting in no subsequent polygon refinement.
The prediction threshold and mask opacity are levers a user can toggle to better capture the object’s pixel area when using one-click SmartPoly.
Demo Enhanced SmartPoly and Image Annotation Tools
Alegion delivers machine learning-powered product features to increase the efficiency and quality of image and video annotation for your machine learning projects. Our software reduces the cognitive load of annotators, letting them focus on more subjective and complex labeling tasks.
Alegion Platform can handle a wide variety of use cases throughout your project lifecycle, from proof of concept, through small batch testing, to full scale production alongside our team of data engineers.