Upload
dorcas-logan
View
221
Download
0
Tags:
Embed Size (px)
Citation preview
Video Matting from Depth MapsJonathan FingerOliver Wang
University of California, Santa Cruz
{jfinger, owang}@soe.ucsc.edu
MotivationGiven a video, replace the background with something differentIsolate the find foreground in each frameGiven a video, replace the background with something differentIsolate the find foreground in each frame
Our MethodUse a depth camera to automate foreground extractionUse Bayesian mattingImprove the matting algorithm to get more realistic video
The Matting ProblemSeparation of a foreground image from a background image
The Easy DirectionBackground(known)Foreground(known)Composite(unknown)2 knowns, 1 unknown
The Hard DirectionBackground(unknown)Foreground(unknown)Composite(known)1 known, 2 unknowns
The Matting ProblemActually there is another unknownRepresents areas that are a combination of foreground and background
The Matting Problem=1=0.5=0Foreground
The Matting ProblemHow do we isolate the foreground?Use an alpha maskAlpha MaskAn image who's color represents foreground and background
The Matting Problemoriginalalpha mask
The Masking ProblemBasic pipelineOriginal compositeAlpha maskIsolated foregroundNew backgroundNew composite
The Masking ProblemBut, how do you get an alpha mask?
Previous WorkBlue Screen MattingPetro Vlahos (1964)Hollywood Special Effects pioneer
Can isolate the foreground if the background is a constant color
Previous WorkBackground is known so it is easy to make a mask
The Matting ProblemHow can this be done with an unknown background?
Use a general matting algorithminput: original composite + trimapoutput: alpha mask
TrimapsA three color image (usually drawn by hand)Black = 100% backgroundWhite = 100% foregroundGray = unknown
TrimapsThe matting algorithm fills in the gray area with estimated alpha values
Matting AlgorithmsThe matting equationFor each 2D location in the image,there is a given composite pixel C
We are to find F, B, and at each pixel whereC = F + (1 - )B
Bayesian MattingOriginal compositeTrimapForeground estimationBackground estimationAlpha mask
Matting Algorithmsalpha maskbackground removedclose upKnockoutRuzon and TomasiBayesian
Problem with Bayesian MattingThese all require a manual trimapOur goal is to do this with videoWe do not want to make trimaps by hand
Previous WorkDefocus Video Matting(McGuire)Two camerasone focused on the backgroundone focused on the foreground
Previous WorkA trimap can be generated from the defocused foregroundHowever, apertures have to be very specific and can be thrown off by lightingAlso requires texture in the scene
Previous WorkBayesian Matting Using Learned Image Priors(Apostoloff, Fitzgibbon)Sequences of frames can be compared in order to find movement
Previous Workassumptionsforeground is movingnothing else is moving
Our ContributionAutomatically generated trimapsDoes not depend on lighting, texture or movementImproved Bayesian Matting using depth informationHella Trimaps
OverviewLow res depthOriginal compositeHigh res depthTrimapNew compositeNew backgroundAlpha maskSupersampleBayesian mattingCompose
Our MethodCanesta depth cameraUses infrared lasers to detect distances from the camera
Our MethodCanesta takes 64x64 resolution imageOptical images are 640x640 or more
Trimap OverviewTo get a trimap Upsample depth image to resolution of optical image Threshold to separate into two colors Erode/dilate to create a gray border around the foreground
UpsamplingUse Qing's supersampled depth methodUse edge cues from high resolution color imageCan increase the depth resolution to up to 100X
ThresholdingAssumptionForeground is in front of backgroundThreshold on a distance planeDone once for entire animation
Erode/DilateGrow unknown area around edges
Improved Bayesian MattingBayesian matting is ill defined when the foreground and background are similar colorsOriginal imageAlpha mask
Improved Bayesian MattingUse depth information in Bayesian Matting optimization stepOriginal imageBayesian mattingDepth map
Improved Bayesian MattingBayesian MattingImproved Bayesian Method
Resultsvideo
ConclusionVideo matting can be done without the user having to manually tweak any individual framesWe were able to improve Bayesian Matting using depth information