Implementation of "Texture Optimization for Example-based Synthesis" by Vivek Kwatra et al.
The output images are generated on 50 iterations.
You can generate a single image:
cmake . && make && ./DisplayImage input/ztext.bmp
Or all of them
./test.sh
- the authors use a memory efficient kd-tree: they only record the position of the source and recover more info as needed.
- we should compute the kdtree only once per scale level, but I ran into odd segfaults errors as I tried...
- we don't compute origin patches for pixels on the edges. Nothing being done, it's very messy... We should get rid of those edge issues with some padding.
- we should remove the OpenNL dependency: our energy minimization consists only of weighted averages...
- This implementation searches for similar patches using color information, not just intensity information.
- Performance needs some tuning: long-distance repetitive structures are not as nicely captured as in the paper. Tuning the initial scale should be enough to fix this.
- Flow-controled synthesis is not covered.
- cmake
- opencv