A chair in a photograph of a living room, for example, can be turned around or even upside down in the photo, displaying sides of the chair that would have been hidden from the camera, yet appearing to be realistic.
This three-dimensional manipulation of objects in a single, two-dimensional photograph is possible because 3D numerical models of many everyday objects - furniture, cookware, cars, clothes, appliances - are readily available online.
’Instead of simply editing ‘what we see’ in the photograph, our goal is to manipulate ‘what we know’
The research team led by Yaser Sheikh, associate research professor of robotics, found they could create realistic edits by fitting these models into the geometry of the photo and then applying colours, textures and lighting consistent with the photo.
In a statement, Natasha Kholgade, a Ph.D. student in the Robotics Institute and lead author of the study said, ‘In the real world, we’re used to handling objects - lifting them, turning them around or knocking them over. We’ve created an environment that gives you that same freedom when editing a photo.’
Kholgade will present the team’s findings on August 13 at the SIGGRAPH 2014 Conference on Computer Graphics and Interactive Techniques in Vancouver, Canada.
Though the system is designed for use with digital imagery, it enables the same 3D manipulation of objects in paintings and in historical photos. Objects that can be manipulated in photos can also be animated; the researchers demonstrated that an origami bird held in a hand can be made to flap its wings and fly away, or a taxi cab shown in a street scene can levitate, flip over to reveal its undercarriage and lift off into the air.
‘Instead of simply editing ‘what we see’ in the photograph, our goal is to manipulate ‘what we know’ about the scene behind the photograph,’ Kholgade said.
Other researchers have used depth-based segmentation to perform viewpoint changes in photos or have used modelling of photographed objects, but neither approach enables hidden areas to be revealed. Another alternative is to insert a new 3D or 3D object into a photo, but those approaches discard information from the original photo regarding lighting and appearance, so the results are less than seamless.
One of the catches to using publicly available 3D models is that the models rarely fit a photo exactly. Variations occur between the models and the physical objects; real-life objects such as seat cushions and backpacks are sometimes deformed as they are used, and appearances may change because of aging, weathering or lighting.
To compensate, the researchers developed a technique to semi-automatically align the model to the geometry of the object in the photo while preserving the symmetries in the object. The system then automatically estimates the environmental illumination and appearance of the hidden parts of the object - the visible side of a seat cushion or of a banana is used to create a plausible appearance for the opposite side. If the photo doesn’t contain pertinent appearance information - such as the underside of a taxicab - the system uses the appearance of the stock 3D model.
Though a wide variety of stock models is available online, models are not available for every object in a photo. But that limitation is likely to subside, particularly as 3D scanning and printing technologies become ubiquitous.
‘The more pressing question will soon be, not whether a particular model exists online, but whether the user can find it,’ Sheikh said. One thrust of future research will thus need to be automating the search for 3D models in a database of millions.
Report finds STEM job candidates facing bias after career break
Can an employer´s preference for a prospective candidate WITH recent experience over one who does not - perhaps through taking a career break - when...