Computer vision AI models rely on correctly labeled data to derive the right object. The challenge of helping verify that the data used for a model is accurate is one that Ann Arbor, Michigan-based startup Voxel51 wants to solve with open-source tools and a commercial service called FiftyOne Teams.
Ann Arbor is home to the University of Michigan, where Jason Corso, co-founder and CEO of Voxel51, works as a professor, and where he got the idea to build the new company. Corso’s research focuses on computer vision applications such as the relationship between video and natural language. In recent years, the adoption of computer vision has grown, as has the size of the datasets.
“When I was a grad I had data sets that ran into the dozens and I could look at every sample,” Corso told VentureBeat. “Now my students came along and they can’t look at a million monsters; it just isn’t possible, so that’s where the need for Voxel51 arose.”
It is a need that has found acceptance in the market and among investors. Today, the company announced it has raised $12.5 million in Series A funding from Drive Capital, Top Harvest and Shasta Ventures, as well as existing investors eLab Ventures and ID Ventures, and the University of Michigan.
The challenge and opportunity of unstructured data for computer vision
Unstructured data takes many forms and includes any type of data that does not fit into a specific data structure format (for example, columns and rows).
One of the most common forms of unstructured data is video content, which is growing exponentially as the number of cameras worldwide continues to grow. Getting value from unstructured video data can be done in several ways. Corso noted that there are technologies that help users extract semantically meaningful information from images, such as simple tools that allow users to search for images taken in a particular location.
While there is no shortage of unstructured image data and large data sets used to train computer vision models, accuracy is a challenge.
“Our whole shtick is that as data sets grew to over 10 million samples, nobody bothered to look at the images anymore,” Corso said.
What Voxel51 does is bridge the gap between what a data engineer does when creating data sets and what that same engineer or his partner does when they train models. The Voxel51 technology supports the visualization of annotations on image data and can be used to identify potential errors and allow users to compare the performance of different models.
Corso explained that Voxel51 allows users to segment data semantically to understand the correctness of a model. For example, through a Python API, a user can query a computer vision dataset to find all images in which one model outperforms another, for images showing a child running down the street.
Open source and the enterprise
Voxel51 started out as an open source product, but alongside the funding announcement, the company is officially launching its FiftyOne Teams enterprise offering, which includes commercial support and additional capabilities.
The open-source project Voxel51 was first launched in August 2020 and has grown over the past two years, with up to 150,000 monthly users. “The open source project is built for a user with local data, with all the data on one system,” Corso said.
In contrast, FiftyOne Teams’ commercially backed offerings include cloud data support, as well as role-based access control (RBAC) to enable multiple users to securely use the same platform. Currently, the commercial service is not offered as a fully managed cloud service, but organizations will still need to run the technology on-premises or in their own cloud instances.
“We envision a future where, at least for certain types of customers, there may be startups that don’t want to deploy locally in their ecosystem, a managed service, but that won’t come out for a while,” Corso said.
The post Voxel51 improves the accuracy of computer vision models and raises $12.5 million appeared first on Venture Beat.