We get general shape information of urban scene as point cloud
using captured images from car-mouted camera,
and make billboard models (fig 1).
Each billboards are placed at center of voxles,
dividing space by the size of the constant and include enogh points (fig 2).
fig 1 : Captured Images and Point Cloud
fig 2 : Relation of Point Cloud and Billboard's Placement
(a) Point Cloud (b) Result of Voxel Dividing (c) Billboard
As shown in Fig 3, each textures mapping billboards change appearance depends on view-direction.
In generally, these textures are redundant because of capturing same object from every direction.
Therefore, we compress each textures by PCA (Principal Component Analysis) algorithm,
reduce input data size.
And also we reconstruct compression textures in parallel, accelerate the processing,
using GPU shader, and implemet the real-time rendering (fig 4).
fig 3 : Textures in Each View-Points
fig 4 : Reconstruction of Textures in GPU
Finally, we show the result of rendering in Fig 5.
fig 5 : Result of Rendering
(a)GroundTruth (b)Rendering Result
Publications
- Yusuke Wakamoto, Hiroshi Kawasaki, Hiroshi Koyasu, Shintaro Ono,
"Photorealistic Wide-Area Rendering System using Micro-Billboard and Compression Texture",
The 18th Annual Conference of the Virtual Reality Society of Japan (VRSJ2013)
- Yusuke Wakamoto, Yasuhiro Akagi, Hiroshi Koyasu, Shintaro Ono, Hiroshi Kawasaki,
"Wide-Area 3D Information Rendering System using Compression Texture and Billboard",
IPSJ SIG-GCAD 2013
|
