We propose a method for automatically guiding patch-based image completion using mid-level structural cues. Our method first estimates planar projection parameters, softly segments the known region into planes, and discovers translational regularity within these planes. This information is then converted into soft constraints for the low-level completion algorithm by defining prior probabilities for patch offsets and transformations. Our method handles multiple planes, and in the absence of any detected planes falls back to a baseline fronto-parallel image completion algorithm. We validate our technique through extensive comparisons with state-of-the-art algorithms on a variety of scenes.
Project page: https://sites.google.com/site/jbhuang0604/publications/struct_completion
15. • Based on non-parametric framework of [Wexler et al. 2007]
16. • Based on non-parametric framework of [Wexler et al. 2007]
• The basic form
17. • Based on non-parametric framework of [Wexler et al. 2007]
• The basic form
• Patch matching cost
• PatchMatch [Barnes et al. 2009]
for efficient nearest neighbor
field search.
19. • Low-level processing is not sufficient
• Idea: guide the completion using mid-level information!
20. • Low-level processing is not sufficient
• Idea: guide the completion using mid-level information!
• Transformation of source patches
21. • Low-level processing is not sufficient
• Idea: guide the completion using mid-level information!
• Transformation of source patches
• Translational
(Photoshop CAF)
22. • Low-level processing is not sufficient
• Idea: guide the completion using mid-level information!
• Transformation of source patches
• Translational
(Photoshop CAF)
• Rotation and scale
(Image Melding)
23. • Low-level processing is not sufficient
• Idea: guide the completion using mid-level information!
• Transformation of source patches
• Translational
(Photoshop CAF)
• Rotation and scale
(Image Melding)
• Projective transform
(Ours)
24. Mid-Level cues for guidance
•Planes:
Cues for how to deform patches
•Regularity:
Cues for where to sample patches
41. Regularity detection – mode detection
• For fronto-parallel scenes, similar to [He and Sun ECCV 2012]
• Can handle perspectively distorted planes and multiple surfaces
44. Guided sampling and propagation
• Slight modification of PatchMatch algorithm
• Plane guided sampling and propagation
• Draw a plane index based on posterior probability
• Draw random samples using rejection-based
sampling
45. Guided sampling and propagation
• Slight modification of PatchMatch algorithm
• Plane guided sampling and propagation
• Draw a plane index based on posterior probability
• Draw random samples using rejection-based
sampling
• Regularity guided sampling
• Randomly draw a offset vector in rectified space
50. Photoshop CAF Statistics of Patch Offsets Image Melding Our Result
[Wexler et al. 2007]
[Barnes et al. 2009]
[He and Sun 2012] [Darabi et al. 2012]
51. Photoshop CAF Statistics of Patch Offsets Image Melding Our Result
[Wexler et al. 2007]
[Barnes et al. 2009]
[He and Sun 2012] [Darabi et al. 2012]
52. Photoshop CAF Statistics of Patch Offsets Image Melding Our Result
[Wexler et al. 2007]
[Barnes et al. 2009]
[He and Sun 2012] [Darabi et al. 2012]
58. Photoshop CAF Statistics of Patch Offsets Image Melding Our Result
[Wexler et al. 2007]
[Barnes et al. 2009]
[He and Sun 2012] [Darabi et al. 2012]
59. What if there are no structures?
• Reduces to baseline image completion algorithms which
use translational patches (same as Photoshop content-
aware fill)
60. What if there are no structures?
• Reduces to baseline image completion algorithms which
use translational patches (same as Photoshop content-
aware fill)
• Works well with man-made scenes
• Backward-compatible with natural scenes
• Tested on 25 natural images from [Kopf et al. 2012]
64. Check out our project website!
• http://bit.ly/planarcompletion
• Full comparison of 110 images
• 85 urban scenes + 25 natural
• Code will be released
soon!
65. Failure modes
• Difficult to find correct demarcation lines (boundary between
regions)
• Need higher-level information
• Missed, false, or inaccurate detection of plane parameters or
regularity
• Computer vision algorithms are not perfect
77. Conclusions
• Mid-level constraints for the win!
• Our work: planes and regularity
• Great for man-made scenes
• Backward-compatible with natural scenes