Date post: | 13-Jan-2016 |
Category: |
Documents |
Upload: | matthew-crawford |
View: | 222 times |
Download: | 0 times |
Frame Decimation Frame Decimation for Structure and for Structure and
MotionMotion
Young Ki BaikYoung Ki BaikCV Lab.CV Lab.
ContentsContents
MotivationMotivation
Frame Selection methodFrame Selection method
Conclusion and Future worksConclusion and Future works
MotivationMotivation
Profit from video sequenceProfit from video sequence CorrespondenceCorrespondence
To save large number of inlier correspondence from small To save large number of inlier correspondence from small motionmotion
X
O
MotivationMotivation
Problem #1Problem #1 Different amount of motionDifferent amount of motion
Fairly small motion to allow automatic matchingFairly small motion to allow automatic matching Significant parallax and large baseline to assure of Significant parallax and large baseline to assure of
a well-condition for 3d reconstruction a well-condition for 3d reconstruction
Uncertainty
boundary
Small base line
Large base line
MotivationMotivation
Problem #2Problem #2 Unsharp frames by Bad focusUnsharp frames by Bad focus
Giving us ill-condition of corner detection.Giving us ill-condition of corner detection. Eliminating unsharp frame to save corner info.Eliminating unsharp frame to save corner info.
MotivationMotivation
Solution → Frame SelectionSolution → Frame Selection Deleting redundant and unsharp framesDeleting redundant and unsharp frames Isotropic motionIsotropic motion
MotivationMotivation
Frame SelectionFrame Selection Frame decimation for structure and motionFrame decimation for structure and motion
David Nister, SMILE2000David Nister, SMILE2000 An assessment of information criteria for motion moAn assessment of information criteria for motion mo
del selectiondel selection Torr, CVPR97Torr, CVPR97
Key frame Selection for Camera Motion and StructurKey frame Selection for Camera Motion and Structure Estimation from Multiple Viewse Estimation from Multiple Views
Thormahlen, ECCV2004Thormahlen, ECCV2004
Frame SelectionFrame Selection
Sharpness measureSharpness measure For removing unsharp framesFor removing unsharp frames
Shot boundary detectionShot boundary detection For selecting subset of framesFor selecting subset of frames
Selection of a subsequence of framesSelection of a subsequence of frames For eliminating redundant frames For eliminating redundant frames
Frame SelectionFrame Selection
Sharpness Measure Sharpness Measure for saving corner informationfor saving corner information The measure of image sharpnessThe measure of image sharpness
The mean square of the horizontal and vertical The mean square of the horizontal and vertical derivatives.derivatives.
22
),(
)1,()1,(),1(),1()(#2
1)(
yxfyxfyxfyxfI
ISIyx
Frame SelectionFrame Selection
Sharpness Measure Sharpness Measure Only to consider the relative sharpness of Only to consider the relative sharpness of
similar imagessimilar images
Sudden
Changes
Frame SelectionFrame Selection
Shot Boundary Detection Shot Boundary Detection Shot boundaryShot boundary
These occur when the camera has been turned off These occur when the camera has been turned off and then turned on again at a new place.and then turned on again at a new place.
Shot boundary
Frame SelectionFrame Selection
Shot Boundary Detection Shot Boundary Detection Shot boundaries are detected by evaluating the correlaShot boundaries are detected by evaluating the correla
tion between adjacent frames after global motion comtion between adjacent frames after global motion compensation.pensation.
Mean of Normalized correlation value (TMean of Normalized correlation value (Tsbsb = 0.75) = 0.75)
Shot boundary detection
0.9
0.5
0.9
valueIntensityI
HomographyH
ROI
IIR
)()()(#
1 2
)(
1112
x
xHx
Frame SelectionFrame Selection
Selection of Subsequence Selection of Subsequence For deleting redundant framesFor deleting redundant frames
Considering two propertiesConsidering two properties Normalized Correlation (NC) constraintNormalized Correlation (NC) constraint Distant constraintDistant constraint
Frame SelectionFrame Selection
Normalized Correlation (NC) constraintNormalized Correlation (NC) constraint Check the mean of NC value between near frames.Check the mean of NC value between near frames. Delete Delete FFii when the mean of NC value between when the mean of NC value between FFi-1i-1 and and
FFi+1i+1 is bigger than T(=0.95). is bigger than T(=0.95).
1iF 1iF
iF
95.0T
Frame SelectionFrame Selection
Distance constraintDistance constraint Check maximum distance in correspondences.Check maximum distance in correspondences. Delete Delete FFii when maximum distance is smaller than when maximum distance is smaller than
TTdd (= image size/10). (= image size/10).
Maximum distance of correspondence
1iF 1iF
iF
Frame SelectionFrame Selection
Selection of SubsequenceSelection of Subsequence Deleting process is repeated until no Deleting process is repeated until no
additional deletions occur.additional deletions occur.
Conclusion and Future Conclusion and Future worksworks
ConclusionConclusion Method of the frame selection in video Method of the frame selection in video
sequencesequence
Future worksFuture works Usage of video sequenceUsage of video sequence ImplementationImplementation