EPFL Spinoff Turns Thousands of 2D Photos into 3D Images

Wouldn’t it be cool if you could upload a bunch of photos of your favorite building and your computer could spit out a stunning 3D model? Researchers in EPFL’s Computer Vision Laboratory thought so. They developed a computer program that generates a 3D image from up to thousands of 2D shots, with all the processing done in the cloud. Since April, the EPFL start-up Pix4D has been offering the modeling service with an intriguing plus: a fourth dimension—time. Now, individuals and small businesses looking for fast, cheap, large-scale 3D models can get them without investing in heavy processing.

“All the information we need to make a 3D model is contained in the photos and the differences we find among them,” explains Pascal Fua, head of the Computer Vision Laboratory. “Along with the image, our program takes additional information such as the time the photo was taken and even corrects the GPS location for more precise localization.”

With Pix4D, users upload a series of photos of an object and within 30 minutes they have a 3D image. The software defines “points of interest” from among the photos, or common points of high contrast pixels. Next, the program pastes the images together seamlessly by matching up the points of interest. Much in the same way our two eyes work together to calculate depth, the software computes the distance and angle between two or more photos and lays the image over the model appropriately, creating a highly accurate 3D model that avoids the time intensive, “point by point” wireframe method.

“For one project, we’ve processed over 50,000 photos, taken mostly by students, and created a high-resolution 3D model of the old city of Lausanne,” explains Pix4D founder and EPFL post-doc Christoph Strecha.

With Pix4D’s 3D models, you can navigate in all directions as well as change the date on a timeline to see what a place looked like at different times of the year. The company is collaborating with several drone makers among which is another EPFL startup called senseFly to market their software as a package with senseFly’s micro aerial vehicles, or autonomous drones.

Adding the fourth dimension
Pix4D’s time element is especially intriguing for senseFly’s customers. Instead of waiting for Google to update its satellite data or for an expensive plane to fly by and take high-resolution photos, farmers, for example, can now send relatively inexpensive flying drones into the air to take pictures as often as they like, allowing them to survey the evolution of their crops over large distances and long periods of time. And since the calculations are done on a cloud server, the client doesn’t need a powerful computer of his or her own.