One of Our Ph.D.’s, Jared Heinly, Presenting This Week at the CVPR Conference

Sneak Peek of the Presentation by Jared Heinly, Ph.D. at the CVPR Conference

Your chance to learn the ins and outs of large-scale 3D reconstruction from a massive volume of images is almost here. For the first time, Dr. Jared Heinly, along with his team, will be presenting a full day tutorial at the CVPR conference that you won’t want to miss.

Dr. Jared Heinly is a Senior Research Engineer with URCV, which has developed the technology behind His Ph.D. research focused on large-scale 3D reconstruction from images. He developed a reconstruction system that processed and reconstructed scenes from a 100 million image dataset within 6 days. This was accomplished using only a single computer— two orders of magnitude in improvement over previously published methods. His research and innovation in 3D software modeling powers the platform, creating the complex and incredibly accurate volume measurements and resulting reports that make a groundbreaking achievement.

Dr. Heinly will be part of a full-day tutorial at the CVPR conference (Computer Vision and Pattern Recognition) hosted in Hawaii in July. CVPR is the premier annual computer vision event, and this tutorial was chosen from many submissions in a competitive process. This tutorial comes on the heels of the very successful event at 3DV, a gathering of the top researchers in the 3D computer vision community in late 2016.

Large-scale image-based 3D modeling has been a major goal of computer vision. The applications are endless, including virtual reality, image-based localization, and autonomous driving and navigation. One of the most diverse data sources for modeling is Internet photo collections – where we are now seeing incredible numbers of photos publicly available, in real-time, on platforms like Twitter and Flickr. With this technology, data modelers can essentially crowdsource and construct real-time models of scenes from around the world.

Bringing a massive collection of a hundred million images together, seeing how the images relate to each other, and fitting them into a 3D construct is no small task. The software must be able to find the relevant pieces of information, see which photos are related and group them together, and filter out those photos which are not useful.

So why is this new method of modeling so significant? “It’s the state of the art, best method out there,” said Dr. Heinly. “It’s tremendously scalable and it makes it very, very easy to process insane amounts of imagery.”

The commercial applications of this technology are very exciting. Already it has been used to construct models of famous buildings from all over the world, using tourist photos as its input. Soon, millions of crowdsourced images, whether taken from autonomous cars or cellphones, will be able to construct real-time models of the world that will give insights we can’t imagine now. Just a few years ago, the vast majority of interest in this field was from academia, but now we can expect a mix of academics and industry professionals driving innovation in the field.

“In the past 5 years, the industry has exploded,” says Dr. Heinly. “There is now a huge commercial interest in large-scale 3D reconstruction.”

Dr. Heinly won’t be conducting the tutorial alone. The team at the conference includes: Dr. Jan-Michael Frahm, Dr. Enrique Dunn, Dr. Marc Pollefeys, and Ph.D. candidate Johannes L. Schönberger. All were affiliated with the University of North Carolina at Chapel Hill and are highly regarded in their field. This tutorial, unlike the shorter one at 3DV, will offer practical insights from the top experts in the field, not just an overview of the technology. Attendees will learn the best methods to use to achieve optimal results, and will have an in-depth look at the algorithms and steps to successful, efficient, large-scale 3D construction. Don’t miss it!