Creating Picture Legends For Group Photos
1 like1,626 views
This document presents a framework developed by Junhong Gao, Seon Joo Kim, and Michael S. Brown for creating picture legends from group photos using automatic human body and face detection techniques. It discusses interactive image segmentation methods and formulates segmentation as a Markov Random Field (MRF) to enhance the accuracy through data terms and smoothness constraints. The authors propose utilizing a head-and-shoulder and upper body prior to simplify the complex multi-labeled segmentation into a binary problem.
Creating Picture Legends For Group Photos
- 1. Creating Picture Legends for Group Photos Junhong Gao† Seon Joo Kim‡ Michael S. Brown† † School of Computing ,National University of Singapore ‡ Department of Computer Science, SUNY Korea
- 4. • Automatic Human Body and Face Detections - Face detection: Viola et al. , Lienhart et al. … - Human body detection: Dalal et al., Zhu et al. … • Interactive Image Segmentation - Boundary based: Kass et al., Mortensen et al. … - Scribble based: Li et al., Rother et al. … - Paint based: Liu et al. …
- 6. 1 2 3 Face Detection Person-wise with Adjustment rectangle cropping 5 Final Legend 4 User Interactive Result Refinement Initial Automatic Segmentation
- 7. 2.5h h w 2w 6h Head and Shoulder Upper Body 4w
- 8. 2.5h h w 2w 6h Head and Shoulder Upper Body 4w
- 9. • Segmentation formulated as a Markov-Random- Field(MRF): • Data term is defined by color mixture model and head- and-shoulder/body prior R G B ×
- 10. • Segmentation formulated as a Markov-Random- Field(MRF): • Data term is defined by color mixture model and head- and-shoulder/body prior • Smoothness is based on intensity difference of adjacent pixels with different labels.
- 13. - - • Over 100 manually segmented training data
- 14. - - • Over 100 manually segmented training data 1.0 0.0
- 15. - - B F H U F B Background region Unknown region Hair region Foreground region
- 16. - + Upper Body Prior
- 17. • Label from starting point of the scribble is set to be foreground L4 label • All the labels that the scribble L3 L1 covered are set to be unknown labels L2 • Other labels are fixed as background labels
- 22. Original Image
- 23. Automatic Segmentation Result
- 25. Final Result
- 27. Photoshop result
- 28. Our result
- 29. • Presented a framework to create picture legends from group photos. • Proposed our head-and-shoulder and upper body prior to assist the segmentation • Simplified a multi-labeled MRF problem into a consecutive binary segmentation problem
- 30. Thank You