Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Video Segmentation April 30 th, 2006
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Introduction Recognition and Segmentation Min Cut Max Flow Single Image Methods –GrabCut –Lazy Snapping –…
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Lazy Snapping Interactive User Interface
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Lazy Snapping Energy minimization
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Lazy Snapping Energy minimization
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Lazy Snapping Boundary overriding
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Lazy Snapping Boundary overriding
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Motivation Obvious Next Step Video Cut & Paste Video Manipulation and Editing
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Introduction Frame by Frame –Time Consuming and Tedious Error With Simple Methods –Fast motions –Deforming silhouettes –Changing topologies
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Introduction Two Papers –Video Object Cut and Paste –Video Cutout
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Video Object Cut and Paste Yin Li, Jian Sun, Heung-Yeung Shum
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Overview
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Pre-segmentation Pre-Segmentation to All Frames
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Key Frames Picking Key Frames.
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Key Frames User Fore/Background Segmentation
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon 3D Graph Cut Segmentation 3D Graph – G=(V,A) Labeling –Foreground = 1 –Background = 0 Volume Between Successive Key Frames
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon 3D Graph Construction 2 Kinds of Arcs: –A I – Intra Frames (BLUE) –A T – Inter Frame (RED)
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon 3D Graph Construction Minimizing Equation: E 1 – Global Color Models E 2 – Penalizing Spatially E 3 – Penalizing Temporally
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Likelihood Energy GMMs Decide Label In Key Frames:
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon GMM Gaussian Mixture Model Distance is Measured By:
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Prior Energies E 2, E 3 Are the Same Distance of Adjacent Regions. β = (2 E (||c r – c s || 2 )) -1
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Prior Energies λ 1 = 24 λ 2 = 12
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon 3D Graph Segmentation
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Errors Global Colors Similarity to Background Thin Areas
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Error Overriding Video tubes Manual corrections
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Video Tubes Local Color Models Put Two Windows Tracking Algorithm Key Frames to Solve W1W1 WTWT
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Fixing graph cut segmentation Minimizing:
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Overriding Brush Fixing Boundary Manually
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Manual error overriding
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Soften hard segmentation Matting
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Coherent Matting Boundary is not 0/1 Prevent Bolting Pixels Smooth Paste
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Coherent Matting
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Example
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Example
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Example
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Example
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Video CutOut J. W ANG, P. B HAT, A. C OLBURN, M. A GRAWALA, M. C OHEN. Interactive Video Cutout. ACM Trans. on Graphics (Proc. of SIGGAPH2005), 2005
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Video Cutout introduction What’s new? Different user interface 3D graph formation Refinement mechanism
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon System overview
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon 3D Graph construction Hierarchical graph nodes: 1.Frame by frame mean shift segmentation 2.Aggregating segments across frames
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Pixel 26-neighborhood induce links Lower level links induce higher level link 3D Graph construction
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Stroking foreground and background over the 3D spatio-temporal volume Not segmenting any frame User Interface
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Graph construction –User input propagates upward –Min cut uses yellow nodes 3D Min cut/Max flow
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon 3D Min cut/Max flow Weights / Energy function –The energy function: –Data term: color similarity to F/B model –Link term: cut likelihood
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon 3D Min cut/Max flow Terms in energy function Graph Cut Energy function L Link D F Foreground D B Background D B,L Pix. historyD B,G ColorD F,G ColorL L Local temporal L G Gradient
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Data weight User input generates color model (GMM) Infinite weight preserves marked pixels Data weight = abiding to F/B color model Graph Cut Energy function L Link D F Foreground D B Background D B,L Pix. history D B,G ColorD F,G Color L L Local temporal L G Gradient
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Data weight White – high probability Foreground Black – Low probability Foreground Graph Cut Energy function L Link D F Foreground D B Background D B,L Pix. historyD B,G Color D F,G Color L L Local temporal L G Gradient
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Data weight White – high probability Background Black – Low probability Background Graph Cut Energy function L Link D F Foreground D B Background D B,L Pix. history D B,G Color D F,G ColorL L Local temporal L G Gradient
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Strong gradients segment border Link cost encourage cut at edges Link weight Graph Cut Energy function L Link D F Foreground D B Background D B,L Pix. historyD B,G ColorD F,G ColorL L Local temporal L G Gradient
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Link weight White – low cut probability Black – high cut probability Graph Cut Energy function L Link D F Foreground D B Background D B,L Pix. historyD B,G ColorD F,G ColorL L Local temporal L G Gradient
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Pixel span: (x o, y o, t) t>0 Data weight Graph Cut Energy function L Link D F Foreground D B Background D B,L Pix. history D B,G ColorD F,G ColorL L Local temporal L G Gradient
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Data weight Graph Cut Energy function L Link D F Foreground D B Background D B,L Pix. history D B,G ColorD F,G ColorL L Local temporal L G Gradient Local background model Assuming camera is stabilized, video is registered Extracting “clean plate” Weight per pixel span
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon d(z i ) = minimum color distance {“clean plate”, B marked pixel}. “Clean plate” cannot be always trusted Weight: Data weight Graph Cut Energy function L Link D F Foreground D B Background D B,L Pix. history D B,G ColorD F,G ColorL L Local temporal L G Gradient
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Data weight White – high probability Background Black – Low probability Background Graph Cut Energy function L Link D F Foreground D B Background D B,L Pix. history D B,G ColorD F,G ColorL L Local temporal L G Gradient
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Link span: links between two adjacent pixel spans Link weight Graph Cut Energy function L Link D F Foreground D B Background D B,L Pix. historyD B,G ColorD F,G Color L L Local temporal L G Gradient
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Strong edges exists within segment Small change over time Local temporal link cost penalize strong temporal gradient Link weight Graph Cut Energy function L Link D F Foreground D B Background D B,L Pix. historyD B,G ColorD F,G Color L L Local temporal L G Gradient
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Link weight Graph Cut Energy function L Link D F Foreground D B Background D B,L Pix. historyD B,G ColorD F,G Color L L Local temporal L G Gradient
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Link weight White – low cut probability Black – high cut probability Graph Cut Energy function L Link D F Foreground D B Background D B,L Pix. historyD B,G ColorD F,G Color L L Local temporal L G Gradient
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Energy function Graph Cut Energy function L Link D F Foreground D B Background D B,L Pix. historyD B,G ColorD F,G ColorL L Local temporal L G Gradient 3D Min cut/Max flow λ2λ2 λ1λ1 λ3λ3
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Iterative process The user refines the cut Adds F/B strokes Graph is re-computed N+1 th iterationN th iteration
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Post Processing
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Post processing Binary cut obtained Edges need refinement
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon A pixel-level min cut around edges Color model obtained form boundary Uniform edge cost = small cut Refinement
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Matting Soften hard segmentation Evaluate α Channel
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Matting Refinement fixed boundary locally Global 3D mesh α Channel along mesh normals
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Results
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Results
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Performance
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Pros –Online 3D min cut –Spatio temporal smooth cut Cons –Does not handle shadows –Ignore motion blur (LPF to avoid temporal aliasing) –Cannot separate translucent objects Summary
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Comparison Video Cutout Video object cut and paste Features 3D segmentation in 2 stages spatial-temporal manipulation 2D segmentation Frame base interface Graph nodes UI Performance 25 min 10 sec per Min cut 30 min 60 min 4-5 min 25 min ? 30 min Preprocessing Artist time Post processing Total
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon Questions?
Advanced Topics in Computer Vision Spring 2006 Video Segmentation Tal Kramer, Shai Bagon The total energy: Foreground and background terms: Background terms: Link terms: Energy function