Presentation is loading. Please wait.

Presentation is loading. Please wait.

Drag-and-drop Pasting By Chui Sung Him, Gary Supervised by Prof. Chi-keung Tang.

Similar presentations


Presentation on theme: "Drag-and-drop Pasting By Chui Sung Him, Gary Supervised by Prof. Chi-keung Tang."— Presentation transcript:

1 Drag-and-drop Pasting By Chui Sung Him, Gary Supervised by Prof. Chi-keung Tang

2 Outline Background Objectives Techniques Results & extended application Demo

3 Background Seamless object cloning Traditional method – User interaction – Time – Expertise

4 Objectives Reduce user-interaction Suppress unnatural look automatically Optimize boundary to achieve the above objectives

5 Techniques User provide rough region of interest (RoI) – Contiaining object of interest (OoI) – Drag-and-drop to the target Optimization problem Euler-Lagrange equation Poisson equation Ω Ω obj f*

6 Problem

7 Objectives Reduce user-interaction Suppress unnatural look automatically Optimize boundary

8 User provides only rough RoI Assume v = ∇ g and let f’=f – g, reformulate optimization problem Poisson equation becomes Laplace equation Approach zero when (f*-g) = constant – find an optimal boundary to satisfy this Techniques (Cont’d)

9 To find the optimal boundary – Inside the RoI – Outside the OoI Define an energy function – Total color variance – Minimize it Ω Ω obj f*

10 Iterative minimization Initialize ∂Ω as boundary of RoI Given new ∂Ω, optimize E w.r.t. k Given new k, optimize E with new ∂Ω – Shortest path problem Until convergence reached

11 Shortest path problem? Cost of each pixel = its color variance w.r.t. new k Path to find in closed band Ω\Ω obj – Not a usual shortest path A shortest closed-path problem Ω Ω obj f*

12 Shortest closed-path Break the band with a cut – Not closed now

13 Shortest closed-path Perform usual shortest path algorithm on a yellow pixel – Dijkstra O(NlogN)

14 Shortest closed-path Perform on M yellow pixels – O(MNlogN)

15 Selecting the cut With minimum length M Reduce probability of twisting path – Not to pass the cut more than once Reduce running time (MNlogN)

16 Results

17

18 Result

19

20 Extended Application Seamless image completion A hole in an image S Another image D provided by user – Semantically correct Auto complete the hole

21 Seamless Image Completion D and S semantically agreed – Color – Scene objects Selecting region on D to complete the hole – Sum of Squared Difference (SSD) of color – Distance to the hole on S

22 Seamless Image completion Result

23

24 Live Demo

25 Q&A

26 THE END


Download ppt "Drag-and-drop Pasting By Chui Sung Him, Gary Supervised by Prof. Chi-keung Tang."

Similar presentations


Ads by Google