Oct 20, 2017 | By Benedict
Researchers from Germany’s Saarland University and Max Planck Institute for Informatics have joined Intel to improve digital 3D object creation using incomplete 3D scanning data. “VConv-DAE” is a convolutional volumetric auto encoder that learns volumetric representation from noisy data.
3D scanning has a wide range of applications, from reverse engineering in the automotive and aerospace sectors, to gathering anatomical data for customized foot orthotics, to simply making 3D printed selfies.
But get yourself an incomplete or distorted 3D scan, and the resulting data can be rendered useless. This isn’t an infrequent occurrence either: improper lighting, movement during 3D scanning, and various other factors can result in badly generated 3D models.
The best way to combat these problems is, of course, to eliminate them at the source, by investing in proper lighting equipment, stable rotating platters, and various other tools. But when those options aren’t available, other routes must be explored.
One potential route has just been developed by a multi-skilled and widely sourced group of researchers, consisting of representatives from computing giant Intel and two German institutes: the Max Planck Institute for Informatics in Munich, and Saarland University in Saarbrücken.
Together, this multidisciplinary team has developed VConv-DAE, a deep volumetric shape learning encoder that learns volumetric representation from noisy data by estimating voxel occupancy grids.
The researchers say the tool is perfect for “challenging tasks like de-noising and shape completion” that can arise during 3D scanning applications.
“Although…3D scanning technology has made significant progress in recent years, it is still a challenge to capture the geometry and shape of a real object digitally and automatically,” says Mario Fritz, leader of the “Scalable Learning and Perception” group at the Max Planck Institute for Informatics.
The kind of 3D scanning equipment Fritz is talking about isn’t necessarily high-end equipment, but things like the Microsoft Kinect, a motion-sensing input device generally used for video gaming on Microsoft Xbox gaming consoles.
One weakness in Kinect-style hardware is an inability to accurately recognize a wide range of textures. This means that surfaces that are too reflective, mottled, or otherwise hard to discern may result in inaccurate 3D data—something that can have knock-on effects for 3D printing.
“The resulting flawed or even incomplete 3D geometries then pose a real problem for a range of applications, for example in virtual or augmented reality, working together with robots, or 3D printing,” Fritz says.
In order to solve these problems, the newly developed VConv-DAE tool uses a special deep learning neural network to generate 3D models from incomplete datasets.
The secret to the encoder’s success, according to the researchers, is avoiding the intuitive mistake of assigning every object with a label: “training as well as the resulting representation are strongly and unnecessarily tied to the notion of object labels,” they say, arguing that their voxel occupancy grid estimation method works much better.
Interestingly, the new technique offers “competitive performance when used for classification,” while also providing “promising results for shape interpolation.”
Ultimately, this could contribute to a new generation of 3D scanning tools that allow simple hardware like the Kinect to produce highly accurate 3D data with no missing information. This, the researchers say, is more a necessity than a hope.
“In the future, it will have to be possible to capture real-world objects simply and quickly, and project them in a realistic way into the digital world,” says Philipp Slusallek, professor of computer graphics at Saarland University and scientific director of the German Research Center for Artificial Intelligence (DFKI).
Slusallek is a leading figure in the European joint research project “Distributed 3D Object Design,” or DISTRO, a network bringing together leading laboratories in Visual Computing and 3D Computer Graphics across Europe “with the goal of training a new generation of scientists, technologists, and entrepreneurs” in the field of distributed 3D object design, customization, and fabrication.
Posted in 3D Scanning
Maybe you also like:
What's Hot Today!