1
naïve composi-ng 2
Poisson composi-ng 3
our result 4
Objec+ves Seamless composi+ng Robust to inaccurate selec+on Output Quality - Limit color bleeding Time- Performance - Efficient method 5
Related Work: Poisson Composi+ng Pas+ng gradients instead of pixels Pérez 03, Georgiev 06, Pros: Blends seamlessly Linear computa+on Cons: Bleeding visible Foreground to background bleeding Poisson Composi+ng Result 6
Related Work: L 1 Norm Introduced in shape from shading Reddy 09 Pros: Reduced bleeding Cons: Nonlinear - Computa*onally intensive L 1 Norm Result 7
Related Work: Moving Boundaries Move the boundaries to avoid bleeding Jia 06 Pros: Avoids bleeding Cons: We don t want boundaries to change Changed composi+on Changing boundaries 8
Contribu+ons Conceal bleeding in textured areas Be_er gradient field at boundary Efficient linear scheme 9
Overview Problem Descrip-on Hiding Residuals with Visual Masking Genera-ng a low- curl boundary Results and Comparisons 10
Foreground Algorithm Overview Selec-on Background
Foreground Algorithm Overview Selec-on grad Background grad
Foreground Algorithm Overview Selec-on grad paste Target gradient field Background grad
Foreground Algorithm Overview Selec-on grad paste Target gradient field Background integrate grad Result 14
Standard Approach: Poisson composi+ng We seek image I with gradients close to target v. target gradient field = 15
Standard Approach: Poisson composi+ng We seek image I with gradients close to target v. target gradient field = Least- squares: output gradient field 16
Our Approach We seek image I with gradients close to target v. target gradient field = value that minimizes curl Weighted least- square: weight output gradient field 17
Overview Problem Statement Hiding Residuals with Visual Masking Genera-ng a low- curl boundary Results and Comparisons 18
Our Strategy Use the weights to locate integra+on residuals where they are less visible Exploit perceptual effect: visual masking human percep+on affected by texture. 19
Visual Masking Demo Can you spot all the dots? 20
Visual Masking Demo Texture hides low- frequency content 21
Visual Masking Demo Can you spot all the dots? 22
Design of the Weights smooth region: needs high weight to prevent bleeding textured regions: low weight is ok because bleeding less visible weights (white is high) 23
Es+ma+ng the Amount of Texture RGB gradient field 24
Es+ma+ng the Amount of Texture Small Gaussian convolu-on RGB gradient field Noise controlling func-on Large Gaussian convolu-on 25
Es+ma+ng the Amount of Texture small Gaussian large Gaussian gradient field g 26
Es+ma+ng the Amount of Texture Texture Map Output (white indicates more texture) 27
Weight Formula = 1 Weight output v depends only on foreground and background does not depend on the unknown I weights are constant in the op+miza+on our energy is a classical least- squares op-miza-on 28
Hiding Residuals with Visual Masking Bleeding only in textured areas Poisson composi+ng With texture weights 29
Hiding Residuals with Visual Masking Residuals Poisson composi+ng With texture weights 30
Hiding Residuals with Visual Masking Reduced bleeding but not fully Poisson composi+ng With texture weights 31
Overview Problem Statement Hiding Residuals with Visual Masking Genera-ng a low- curl boundary Results and Comparisons 32
Boundary Problems Target field v integrable iff curl(v)=0 Only boundary has non- zero curl Consequence of target field construc+on (see paper) curl using standard strategy Our strategy: reducing the curl 33
Naïve Approach Minimize Unfortunately, the result is not seamless: output of naïve approach close- up 34
Our Approach: Least Squares Trade- off Unknowns: target field v along boundary Weights depend on texture (detail in paper) zero curl weight seamless composi+ng 35
Overview Problem Statement Hiding Residuals with Visual Masking Genera-ng a low- curl boundary Results and Comparisons 36
Results and Comparisons Copy- and- paste 37
Results and Comparisons Poisson Reconstruc+on : Pérez 03, Georgiev 06, 38
Results and Comparisons Maximum Gradient : Pérez 03 39
Results and Comparisons Diffusion : Agrawal 06 40
Results and Comparisons Lalonde 07 41
Results and Comparisons L 1 Norm : Reddy 09 42
Results and Comparisons Our Result 43
Results and Comparisons Our Result L_1 Norm Lalonde Diffusion Max Poisson 0 5 10 15 20 Running Time (Seconds) 44
Results and Comparisons Poisson 45
Results and Comparisons Our Result 46
Results and Comparisons Poisson 47
Results and Comparisons Our Result 48
Discussion Several parameters - all examples use the same parameters Discolora+on may happen - happens in all gradient based operators Our model of visual masking is simple - good for performance - more complex model could be used 49
Conclusion Robust composi+ng using visual masking Be_er gradient field at boundary Efficient linear scheme Poisson Reconstruc+on Our Result 50
Special Thanks Todor Georgiev Kavita Bala George Dre_akis Aseem Agarwala Bill Freeman Tim Cho Biliana Kaneva Medhat H. Ibrahim Ravi Ramamoorthi This material is based upon work supported by the Na+onal Science Founda+on under Grant No. 0739255 and No. 0924968. 51
Conclusion Robust composi+ng using visual masking Be_er gradient field at boundary Efficient linear scheme Poisson Reconstruc+on Our Result 52
Results and Comparisons 53
Results and Comparisons Poisson 54
Results and Comparisons Our Result 55