European Conference on Computer Vision 2014

Real-Time Exemplar-Based Face Sketch Synthesis

Yibing Song1     Linchao Bao1     Qingxiong Yang1     Ming-Hsuan Yang2    



Topic Background


Sketch synthesis is trying to simulate the hand drawing style of artist on generating new images automatically. Existing methods like PhotoShop produce sketch effect according to image edges, which is suitable for normal images. However these methods commonly fail to capture important facial details, which would be sensitive to human beings. So the generated sketches are more like the input images rather than artist work. To solve this problem a training dataset is involved which contains photo-sketch pairs. The output sketch is reconstructed from training sketches through mapping functions learned from input and training photos.



Spotlight


Prior art methods : Sparse patches represent face photo + KNN search for sketch patch selection + Global sketch patch optimization using Markov Random Field.
Proposed : Dense patches represent face photo + KNN search for sketch patch selection + Spatial / Temporal sketch denoising.

Contributions:
1: More efficient than prior art on CPU.
2: The first method to enable real-time performance on GPU.
3: The first method to enable temporal coherence for video sketch synthesis.





Downloads


[Slides.pptx] : Slides containing animation are provided to illustrate the proposed framework
[Paper.pdf] : The paper
[Code.zip] : A self-contained C++ implementation in VS2012. Training dataset is also included. We also provide our implementation of MRF sketch synthesis here.
[Demo.zip] : CPU based webcam input live face sketch synthesis plus color rendering. You can try it on your side.
[Results.zip] : Complete result of state-of-the-art methods on the benchmark dataset
[Poster.jpg] : The poster


BibTex (DOI)


@inproceedings{song_eccv14_sketch,
   author = {Song, Yibing and Bao, Linchao and Yang, Qingxiong and Yang, Ming-Hsuan},
   title = {Real-Time Exemplar-Based Face Sketch Synthesis},
   booktitle = {Proceedings of European Conference on Computer Vision},
   pages={800-813},
   year = {2014},
   }



Last update: 08/10/2014