1. Video Motion Interpolation for Special Effect Applications Timothy K. Shih, Senior Member, IEEE, Nick C. Tang, Joseph C. Tsai, and Jenq-Neng Hwang, Fellow, IEEE IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS— PART C: APPLICATIONS AND REVIEWS, VOL. 41, NO. 5, SEPTEMBER 2011

2. Outline Introduction System Overview Motion Layer Segmentation and Tracking Motion Interpolation Using Video Inpainting Experimental Results Conclusion

3. Introduction

4. Background Video forgery (video falsifying): A technique for generating fake videos by altering, combining, or creating new video contents For instance, the outcome of a 100 m race in the olympic game is changed.

5. Introduction Example of video forgery : Original video frame Falsifying result

6. Objective To create a forged video, which is almost indistinguishable from the original video To create special effects in video editing applications

7. Introduction To change the content of video, the following techniques are commonly used: object tracking motion interpolation video inpainting video layer fusing

8. Introduction Contributions of this paper: 1) It is the first time that video forgery is attempted based on video inpainting techniques. 2) A new concept called guided inpainting for motion interpolation of video objects is proposed. 3) A guided quasi-3-D (i.e., X, Y, and time) video inpainting mechanism is proposed.

9. System Overview

11. 1) Motion Layer Segmentation: Separates background and tracked object 2) Motion Prediction: Finds Reference Stick-Figure to predict cycle of motion 3) Motion Interpolation: Motion analysis Patch assertion Motion completion via inpainting.

12. System Overview 4) Background Inpainting: Inpaints background of different camera motions 5) Layer Fusion: Merges an object layer and a background layer

13. Motion Layer Segmentation and Tracking

15. Separate target objects from the background Adopt Mean Shift Feature Space Analysis Algorithm [2] for color region segmentation [2] D. Comaniciu and P.Meer, “Mean shift: A robust approach toward feature space analysis,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 24, no. 5, pp. 603– 619, May 2002.

16. Initial segmentation of objects from their background

17. ALGORITHM: REFERENCE STICK FIGURE TRACKING Manually Selected Mean Shift Algorithm C0C0 C0’C0’ Frame 0

18. ALGORITHM: REFERENCE STICK FIGURE TRACKING Revised Fast Tracking Mechanism [6] Bounding Box B 1 Frame 1 [6] K. Hariharakrishnan and D. Schonfeld, “Fast object tracking using adaptive blockmatching,” IEEE Trans.Multimedia, vol. 7, no. 5, pp. 853–859, Oct. 2005.

19. ALGORITHM: REFERENCE STICK FIGURE TRACKING Mean Shift Algorithm B1’B1’ Comparing Color Segments of C 0 ’ and B 1 ’ C1’C1’ Frame 1 Frame 0

20. ALGORITHM: REFERENCE STICK FIGURE TRACKING Applying Dilation C1’C1’ C1*C1* Frame 1

21. ALGORITHM: REFERENCE STICK FIGURE TRACKING Frame 0 CoCo C1*C1* Comparing corresponding pixel p If (p in C 1 *) - (p in C 0 ) > L 2 Exclude p in C 1 ’ Else Keep p in C 1 ’ Frame 1 ‧ Set L 2 = 2 ‧ Using L in LUV color space

22. Motion Segmentation Different parts of the target may move in different directions Decomposing an object into different regions Using revised block searching algorithm [10] to compute motion map [10] J. Jia, Y.-W. Tai, T.-P.Wu, and C.-K. Tang, “Video repairing under variable illumination using cyclic motions,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 28, no. 5, pp. 832–839, May 2006.

23. Motion Segmentation The Mean Shift color segmentation can also be revised to deal with motion segmentation: Based on blocks, not pixels Important for video inpainting Ghost shadows can be eliminated

24. Original Video Frame Corresponding result of color segmentation by using [2] The example of tracked object and estimated vectors

25. Motion Interpolation Using Video Inpainting

27. Motion Interpolation : Motions of the target object need to be interpolated. Video Inpainting : In order to obtain the interpolated figures Motion interpolation may create background holes.

28. Motion Interpolation of Target Objects A target object can be segmented into a layer. Motion interpolation is required to produce a slow motion of the target layer. Original Interpolated tntn t n+1 t n+2 t n+3

29. General Inpainting Strategy In order to obtain the interpolated figures Using a rule-based thinning algorithm [1] : To obain the stick figures of target objects Stick figures: Used to guide the selection of patches Copied from the original video [1] M. Ahmed and R. Ward, “A rotation invariant rule-based thinning algorithm for character recognition,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 24, no. 12, pp. 1672–1678, Dec. 2002.

30. General Inpainting Strategy Quasi-3-D video space (2-D plus tIme) Using 3-D patches in quasi-3-D video inpainting produce a smooth movement

31. Prediction and Interpolation of Cyclic Motion Consider the following scenario: 1) It’s common for target objects to perform actions in a repeated cycle. 2) A stick figure can be used to estimate the relative positions of patches. (e.g., head, body, and legs).

32. Prediction and Interpolation of Cyclic Motion Stick figures and the contours of target objects can be used to predict repeated cycles.

33. Prediction and Interpolation of Cyclic Motion Missing stick figure can be reproduced by: 1) Searching for similar reference stick figures in a repeated motion cycle 2) Interpolation of two known stick figures

34. ALGORITHM: REFERENCE STICK FIGURE SEARCHING

35. x …… x+r r : the number of frames in a repeated cycle index range function of a given frame number x as idx(x) = [x + r − 2, x + r − 1, x + r, x + r + 1, x + r + 2] ∪ [x − r − 2, x − r − 1, x − r, x − r + 1, x − r + 2]. x-r

36. STICK FIGURE INTERPOLATION OaOa Union of O a and O b ObOb Thinning result

37. Motion Interpolation Alogorithm extending our image inpainting algorithm for motion interpolation consider a video as a 2-D plus time domain

38. Motion Interpolation Alogorithm I 3 = Φ 3 ∪ Ω 3 Φ 3 is a source space Ω 3 is a target space Φ 3 ∩ Ω 3 = ∅ (an empty set)

39. ALGORITHM: PATCH ASSERTION

40. Example for patch assertion: ALGORITHM: PATCH ASSERTION Patches on stick figure Contour ω of (a) (searched in the nearby motion cycle) Result of patch assertion

41. The main algorithm Let ∂Ω 3 be: a front surface on Ω 3 adjacent to Φ 3 ALGORITHM: MOTION INTERPOLATION 3 3 3 3 Source Region Target Region

42. Given a 3-D patch Ψ p centered at the point p Let Ψ p ‘s priority P(p) = C(p) × D(p) ALGORITHM: MOTION INTERPOLATION 3 3 3 3 3 3 Source Region Target Region

43. C(p) : Confidence term: The percentage of useful information inside a patch centered at p the size of 3-D patch is denoted as |Ψ 3 | = 27 pixels ALGORITHM: MOTION INTERPOLATION

44. D(p) : Data term Compute the percentage of edge pixels in the patch ( Instead of computing the isophote[3] ) var(Ψ p ) : the color variation of the patch [21] ALGORITHM: MOTION INTERPOLATION 3 [3] A. Criminisi, P. Perez, and K. Toyama, “Region filling and object removal by exemplar-based image inpainting,” IEEE Trans. Image Process., vol. 13, no. 9, pp. 1200–1212, Sep. 2004. [21] T. K. Shih, N. C. Tang,W.-S. Yeh, T.-J. Chen, andW. Lee, “Video inpainting and implant via diversified temporal continuations,” in Proc. 2006 ACM Multimedia Conf., Santa Barbara, CA, Oct. 23–27, 2006, pp. 133–136.

45. Example of motion interpolation via inpainting ALGORITHM: MOTION INTERPOLATION

47. Inpainting Camera Motions Using mechanism proposed in [22] Ensures that there is no “ghost shadows” created in the background Segment motions into different regions The inpainted area in the previous frame needs to be incorporated. [22] T. K. Shih, N. C. Tang, and J.-N. Hwang, “Ghost shadow removal in multilayered video inpainting,” in proc. IEEE 2007 Int. Conf.Multimedia Expo, Beijing, China, Jul. 2–5, pp. 1471–1474.

49. Layer Fusion Need to merge video layers to produce forged video. The fusion process merges an object layer and a background layer. (With contour of object layer computed based on the object tracking)

50. ALGORITHM: LAYER FUSION

51. Object layer obj Dilation area of obj δ ob j Corresponding area on background layer δ bkg δ ob j pixel intensity → υ L ObjectBackground

52. ALGORITHM: LAYER FUSION (-2,70) Histogram of the intensity difference in δ ob j and δ bkg υLυL Intensity Difference Count

53. Experimental Results

54. Block size used in Patch Assertion: 10-by-10 pixels Patch size used in Motion Interpolation: 3-by-3-by-3 pixels Hardware: CPU 2.1G with 2G RAM

55. Evaluation Motion Layer Segmentation Motion Prediction Motion Interpolation Background Inpainting Layer Fusion

56. Evaluation

60. Limitations of the Proposed Mechanism Shadows cannot be tracked precisely. Only use intensity to merge layers (without considering the chrominance information) A sophisticated 3-D reconstruction mechanism needs to be investigated. Only produces actions two times slower for slow motion.

61. Conclusion

62. This paper Proposes an interesting technology to alter the behavior of moving objects in a video Effectively extends the inpainting technique to a quasi-3-D space Allows a video to be separated into several layers, played in different speeds, and then merged

63. Conclusion A series of difficult problems are solved. Solutions are successfully integrated. Mostly, it is a subjective feeling of how a fake video looks real. Necessary to develop an authoring tool to allow the users to specify the spatiotemporal fluctuation property