A group of MIT researchers have developed a method to generate 3D objects that can fool Google’s InceptionV3 neural network image classifier. Google’s InceptionV3 is a deep learning algorithm used to train computer vision. Make sense? For a very basic example, computer vision enables Facebook and similar apps to suggest who is in a picture and will enable a car to drive itself while you Facebook.
To carry out their research under a tight deadline, the team used ZVerse’s LAYR™ software to work with 3D printing technology and interact with 3D experts in real time. LAYR™ is a 3D Design On Demand platform, used in On Demand Manufacturing, providing the fastest path from idea to physical object. With this capacity, the team produced 3D objects that could be carefully colored at high DPI; “The only option we had was 3D printing – anything else would be too expensive or have too high turnaround time”- Logan Engstrom.
Because of the project’s trial-and-error nature, the team needed a solution that they could directly collaborate with experts, and LAYR™ gave them that capacity. The ability to direct message 3D specialists and designers in real time was “invaluable to [their] research effort,” (L.E.) and the ease of sharing large files, requesting quotes, and receiving real-time feedback on the LAYR™ platform expedited their research exponentially.
“None of us were experts in 3D printing, and we knew nothing about the process or constraints on 3D color printing in sandstone. We could not have been brought up to speed without talking to ZVerse experts over LAYR™, and without their advice and tips we could not have printed models that were high quality but inexpensive. LAYR™ made the whole process relatively easy and straightforward, even though we knew close to nothing about 3D printing.”
– Logan Engstrom
Our collaboration with the MIT team fostered discoveries in Three-Dimensional Image Misclassification that could not have been possible without our LAYR™ 3D Design on Demand platform. In particular, the team printed 3D objects (turtles, barrels, oranges, and baseballs) that strongly classified as a desired target item (rifles, guillotines, power drills, and espresso, respectively) over various angles, viewpoints, and lighting conditions by a standard ImageNet classifier. Read more about their discoveries here.
“This MIT research project is a great demonstration of how the ZVerse LAYR™ software bridges the gap from ideation to physical part with incredible efficiency, so innovators can do what they do best. LAYR™ sits at the intersection of 3D Design, On Demand Manufacturing, and customers who need a way to painlessly create manufacturable files.”
–ZVerse CEO John Carrington
With ZVerse’s LAYR™ platform, we are lifting the barriers of 3D Printing, and bringing its capabilities to the curious.