Google has created a solution to speed up annotations for artificial intelligence. This is Fluid Annotation, which uses machine learning to annotate labels for classes, each object and background area in a photo. That’s what the company writes in an announcement.
Annotation is a time-consuming process, especially in the case of images. At this moment every object in an image is still annotated by people, which can sometimes take up to 19 minutes per photo. A complete dataset of 164,000 images can therefore require 53,000 hours of work.
Fluid Annotation replaces that human work with machine learning. According to Google itself, this allows it to accelerate the creation of labelled datasets three times over.
Fluid Annotation starts with the output of a previously trained semantic segmentation model, which generates approximately 1,000 image segments with class labels and reliability scores. The segment with the highest reliability is forwarded to human workers for labelling.
Annotators can also customize a photo via a dashboard, choosing what to correct and in what order. They can also swap a label of an existing segment with another label of an automatically generated shortlist. It is also possible to add a segment to hide a missing object, delete an existing segment and change the depth order of overlapping segments.
“Fluid Annotation is a first step in making annotations for images easier and faster,” said Jasper Uijlings and Vittario Ferrari, Google researchers. “In the future, we want to improve the annotation of object boundaries, make the interface faster by adding more machine intelligence, and eventually expand the interface to handle previously missed classes.
However, Google is not the only one using AI to annotate data. The startup Scale uses a combination of human data labelers and machine learning algorithms to sort raw, unlabeled streams for customers such as Lyft, General Motors, Zoox, Voyage, nuTonomy and Embark.This news article was automatically translated from Dutch to give Techzine.eu a head start. All news articles after September 1, 2019 are written in native English and NOT translated. All our background stories are written in native English as well. For more information read our launch article.