AI contests often ask participants to submit a model, reinforcing the idea that better code is the answer. Pioneering researcher Andrew Ng last fall flipped that notion on its head, asking contestants for a better dataset.
Innotescus, based in Pittsburgh, came within a hair of the top spot in Ng’s Data-Centric AI Competition, applying its data-focused chops.
The company offers a platform for annotating, analyzing and curating images and videos to help developers create better datasets for improved model accuracy.
Its automated NVIDIA GPU-driven video annotation feature, dubbed AVA, enables customers to spend less time manually labeling video frames. Using AVA, annotators can create a single label of an image in videos, and then AVA tracks it through successive frames to automatically annotate.
Innotescus is a spinout from ChemImage, a chemical imaging software specialist. ChemImage was initially using the team’s work to build and deploy models in the life sciences industry.
“Innotescus has allowed us to build the tools we wish we had access to when we faced these same challenges ourselves,” said Chris Anderson, formerly of ChemImage, who co-founded and serves as CTO of the startup.
NVIDIA TAO Toolkit Acceleration
Innotescus is a member of NVIDIA Inception, a program that offers go-to-market support, expertise and technology for AI, data science and HPC startups.
NVIDIA’s developers relations team guided Innotescus on using NVIDIA GPUs as well as the NVIDIA TAO Toolkit to accelerate its work on data annotation.
TAO provides Innotescus access to transfer learning to improve annotation models for improved preprocessing of datasets.
“NVIDIA TAO toolkit provides Innotescus users access to powerful transfer learning techniques to tailor models for custom use cases with less data,” said Anderson.
Better Inventory Keeper
Innotescus assists Bossa Nova Robotics with retail inventories.
Bossa Nova develops a retail AI engine for helping customers apply image-based analytics to inventory. The system is capable of recognizing more than 300,000 unique products, relying on 134 million labeled images from different positions and lighting situations.
But the company has much more data to collect and label for retailers. Yet their home-brewed platform, based on a popular open source one, was a “burden to maintain” and “difficult to teach new users” on, according to Bossa Nova.
Tapping into Innotescus, Bossa Nova has made a threefold increase in its productivity, according to the retail AI company.
Smarter Robot Brains
Pittsburgh-based RE2 Robotics, a Carnegie Mellon spinout that develops mobile robotic arms for outdoor applications, needed more high-quality data to improve its perception models. But with limited budget and time, it didn’t make sense for RE2 to build its own annotation tool or customize an open source one.
So, it turned to Innotescus’s AVA and the results have been huge gains in its model and productivity.
In just more than a month, RE2’s two annotators had manually annotated less than 2,000 images. But using AVA’s object-tracking algorithm on its video annotation, RE2 was additionally able to process more than 40,000 annotated images during the period.
“RE2 was able to reap significant time savings from the decision to leverage the power of AVA,” the robotics company said.
The added data also enabled RE2 to improve its inference on a particular class of images to 90 percent from 50 percent.