“Populating the virtual world with animals used to be a lot of manual work“, says Michael J. Black, director of the Perceiving Systems department at the MPI for Intelligent Systems. Today, animators have to laboriously design 3D animals using graphics software. While there are technologies to create realistic 3D avatars of humans, no such technology is available to automatically capture the shape of wild animals. “You can’t bring a tiger into the lab to be scanned and you can’t take expensive scanners out into nature,” says Black.
The key is to start with a rough 3D shape
The team, whose technique creates avatars of animals including their fur or skin with reasonable effort, includes scientists from the CNR Institute for Applied Mathematics and Information Technology (IMATI) in Milan and the University of California in Berkeley as well as Max Planck researchers. "Ultimately, 3D animal capture is about reconstructing the 3D structure from 2D images of moving objects captured by various unknown and moving cameras," explains Silvia Zuffi, the lead author on the project from IMATI.
In their new method, the researchers only need a few photos of an animal in different views to calculate realistic models that can then be animated naturally. Unlike other approaches, the images from which the avatars are generatedcan also come from different cameras. The computer program delivers such convincing results because it doesn't start from scratch when it creates a model. “The key is to start with a rough 3D shape model that can explain the shapes of many animals and then refine this based on the observed images,” says Michael J. Black.
Helping endangered animals in the wild
In their current work the researchers demonstrate that their method works for as different animals as a tiger, a bear, a horse or a rhino. "Our algorithm generates realistic models of animals for which no previous 3D models exist," says Silvia Zuffi. For example, the researchers were able to reconstruct the 3D shape of the extinct Tasmanian tiger from low-quality film footage shot in the 1930’s. The researchers now want to make the method fully automated so that it can be run on thousands of images of animals in the wild.
Fully automated 3D animal shape capture would not only facilitate the work of game developers and animators but also supports animal conservation. Given photographs of animals captured by camera traps, the method could determine whether their shape, i.e. their weight, has changed. Species conservationists would thus receive quick and reliable statements as to whether endangered animals are doing well.
PH / MJB
The work appears in the annual IEEE Conference on Computer Vision and Pattern Recognition (CVPR) held in Salt Lake City. CVPR is highly competitive and is regarded as the world´s top conference in the field.
The work appears as:
Silvia Zuffi, Angjoo Kanazawa, Michael J. Black, “Lions and Tigers and Bears: Capturing Non-Rigid, 3D, Articulated Shape from Images,” Proc. CVPR, 2018.
For more about the research see: smalr
The 3D models are freely available for download from: sketchfab