The Dataset that Made the Cover

February 27, 2018

By Meredith Alexander Kunz, Adobe Research

When Qingnan Zhou, a research engineer at Adobe Research, and his collaborator, Alec Jacobson of the University of Toronto, first developed their 3D dataset, it was to “stress test” a new algorithm. Could their tool perform correctly on 10,000 different 3D models?

They were excited to discover it worked. To show just how many models they’d covered, the team created a huge poster-style, full-color image including as many of the objects as they could pack in. Now, they are even more thrilled that the image they created from the dataset—already an award winner in 2017—is being featured as the cover of Computer Graphics Forum for February 2018.

“The idea is that when people see this image, they say, ‘Wow, there are a lot of 3D models!’ And if you look in more detail, you see a wide range of geometries and shapes. Some organic, some mechanical, some sculptural,” says Zhou.

The dataset, known as Thingi10K, analyzes models uploaded on the 3D design-sharing site Thingiverse between Sept. 16, 2009 and Nov. 15, 2015. Prior to winning the journal cover contest, it received the 2017 Eurographics Symposium on Geometry Processing Dataset Award.

In the future, this work could have new applications, perhaps serving as a machine learning training set, or helping build new kinds of 3D models.

“It is still very early for 3D geometry, and people are exploring,” Zhou explains.

Zhou didn’t stop there. He decided to create a version of the dataset image that could challenge him in a whole new way: “The image is now an insanely hard puzzle,” says Zhou. His colleagues at Adobe Research used an online service to order the tricky brainteaser, which now is stumping people at the team’s San Francisco office.