Dr Tan Ping showing the 3D model of the Pavilion of Manifest Benevolence (in the Forbidden City, Beijing) reconstructed with a single input image.
THE saying goes that Rome was not built in a day. But it may be possible with Dr Tan Ping's new algorithm. Well, perhaps not in a day but the effort will definitely be a breeze compared to current methods available for reconstruction of 3D models.
Dr Tan, who is with the Department of Electrical and Computer Engineering, has come out with a new algorithm that helps you reconstruct buildings in 3D using a single image. With this new method, ancient Rome can be reconstructed quickly. Furthermore, the reconstructed models using this new method are highly realistic models, sporting fine details such as wall texture and other intricate architectural designs.
Previously, such high quality 3D models were reconstructed rather painstakingly. Such efforts require manual methods which make use of blue prints as guide. Many historic and traditional buildings have intricate geometric structure and curved roofs which are highly non-planar. Though manual methods can capture such details, current image-based techniques are inadequate for the job.
A paper on the novel method was presented by Dr Tan and his team mates, Assoc Prof Cheong Loong-Fah and graduate student Jiang Nianjuan at the SIGGRAPH Asia Conference at Yokohama, Japan in December 2009. The paper was also published in the ACM Transactions on Graphics (TOG). Entitled "Symmetric Architecture Modeling with a Single Image", the paper was stringently reviewed and selected from 275 submissions.
Explaining the applications this new method of constructing 3D models, Dr Tan said, "Sometimes, especially for historic buildings, only one image is available. But from this one image, we may be able to reconstruct the entire building, offering historians, archeologists and architects, insights of the original building from many perspectives."
Their method, said Dr Tan, exploits symmetries for 3D reconstruction. He and his team were inspired by architecture which makes use of symmetry as a unifying factor. Since symmetry is basically a repetition of identical patterns, it effectively upgrades a single input image to multiple images. Using the new toolkit, one can delineate key parts from the image with a few manual strokes, to build an initial model on his screen.
"Our method does not require complicated geometric and photometric image alignments. Our method can be a standalone toolkit for artists to create 3D architecture models from online pictures or archived pictures," Dr Tan elaborated.
Their technique which was developed from a Microsoft grant would also be useful for graphics in computer games as well as film animation.
Looking forward, Dr Tan hopes to fine-tune the algorithm they have developed, making it even more user friendly. He also hopes to create a 3D reconstruction of the entire NUS campus.