mobile off
ABOUT US  |  RESEARCH  |  PUBLICATION   |  COURSE  |  SEMINAR  |  LINKS
Research
Physically-Based Simulation
Fluid Interaction
Fire
Ice and Snow
Water
Smoke
Shell Deformation
User Behavior Analysis
Sentiment Analysis
Outlier Detection
Automatic Player Behavior Analysis
Adaptive Agent Navigation
Crowd Simulation
Graphics Applications
Neuroimage
Artificial Life
3D Geometry Processing
Real domain data visualization
Integral MLS Surface Model
SDF-based Geometry Synthesis
Point-based Geometry and Modeling
Texture Processing
SDF-based Geometry Synthesis

1. Background

Modern computer games and movies often involve complicated large-scale scenes, such as streets of buildings, outback scenes, and caves. Creating large and detailed scenes manually is laborious, so the automatic synthesis of scenes from a small exemplar will be helpful. Recent advances in example-based image and texture editing techniques have provided novice users the ability to manipulate 2D images easily. But the example based synthesis of 3D models is less well explored. We expand this 2D approach into the 3D SDF in this study [1].

fig1
Fig. 1. Adaptively synthesized signed-distance field models. The original signed-distance field model (top left) is synthesized with the proposed method to generate various versions (others).


 
 
2. Technical Details

A. Synthesis Framework

The overall synthesis process basically consists of an upsample-jitter-correction loop, which is repeated until the texture is synthesized in satisfactory detail. Algorithm 3.1 summarizes our synthesis frame work:

fig1

B. Adaptive Offset Map (AOM)

Our synthesis map S and a k-coherence candidate map C represent coordinates in the exemplar E. These maps are spatially coherent inside patch boundaries and should be represented by an easily updatable data structure. Existing octree data representations efficiently represent spatial data. But there is no explicit spatial coherence inside an octree cell, so updating it while retaining subcell coherence is complicated. We propose an adaptive offset map to store spatially coherent offset values. Basically, the AOM is a kind of octree and each cell c (i.e., a node of the octree) stores a vector which maps to one location on the exemplar from the center of the cell.

fig1
Algorithm 4.1 is the pseudocode for evaluating AOM:


 
 
3. Results



fig1

Our method progresses from coarse to fine scales, and the quality of the synthesis improves incrementally. Fig. 12 shows how the resulting scene becomes more coherent as the synthesis map S is incrementally refined at each level. Refinement of the scale improves the visual coherence, because its neighborhoods become more similar to those in the exemplar.


 

Fig. 15. Large-scale scene generated by initializing S with deterministic noise at a coarsest level.


 
 
4. References

[1] Sung-Ho Lee, Taejung Park, Jong-Hyeon Kim and Chang Hun Kim, "Adaptive Synthesis of Distance Fields," IEEE Transactions on Visualization and Computer Graphics, Volume 17, Number 7, pp. 1202-1206, July 2012.


ABOUT_US  |  RESEARCH  |  PUBLICATION  |  COURSE  |  SEMINAR  |  LINKS