One question we get asked a lot at Alegion is how we leverage machine learning to optimize annotation efficiency and quality, and do it at scale. Not surprisingly, the answer is, “It depends.” We invest strategically in ML in a way that allows us to deploy it tactically, and tailored to the specifics of the use case, the scale, and the length of the project.
But, in general, we deploy ML in a few distinct ways:
- Pre-labeling: Giving workers pre-labeled tasks that simply need review and refinement
- Work deflection: Using high confidence model predictions to route easy cases away from human annotation and quality assurance entirely
- Worker scoring: Statistical and ML techniques that provide confidence values for worker skills based on past performance
- Worker augmentation: Giving workers in task ML and CV powered tools that assist in repetitive tasks, reduce cognitive load, and allow workers to focus on task elements that benefit more from human judgement
One of the latest additions to the Alegion Platform, SmartPoly, falls into the last category. SmartPoly takes the extreme top, left, right, and bottom points of an object in image or video as input from the annotator and uses a convolutional neural network (CNN) to segment the object. SmartPoly then renders a polygon based on the segmentation that the annotator can refine as needed. Since the model is classless, and does not use object detection, it can be deployed, without training, in a wide variety of use cases. SmartPoly ML augments the annotator’s segmentation abilities. A bionic arm of sorts, for the annotator.
The time to return a polygon for instance segmentation is typically about .5 seconds and the accuracy is high enough that the polygons returned usually need little or no refinement from the annotator. In a recent complex instance segmentation task involving automotive parts and damage, SmartPoly yielded a reduction in time on task of about 30%. Even better, it requires zero configuration or model training.
Because this type of ML delivers efficiency in a wide variety of use cases with zero setup, we see gains throughout the project lifecycle - from proof of concept, through small batch testing, and full scale production. These tools also keep delivering gains alongside other ML methods we deploy as a project scales.
SmartPoly is easy to use and does the heavy lifting - reducing the cognitive load of annotators and letting them focus on the more subjective and higher order elements of a task. After all, who wouldn’t want to have a bionic arm?