Structure from MotionRoadmapStructure and Motion spaceStructure Estimation - BackgroundPriniciple of FactorizationSlide 6Kanade-Lucas-Tomasi (KLT) trackerSlide 8Slide 9Affine CameraSlide 11Rigid Body FactorizationRigid Body Factorization (contd.)Experiments – Rigid Body FactorizationSlide 15Slide 16Rigid Multi body FactorizationShape Interaction MatrixMulti-body factorization algorithmNon-Rigid 3D Shapes from Image StreamsNon-rigid Body FactorizationSlide 22Slide 23Rank-Constrained Tracking(1)Rank-Constrained Tracking(2)Rank-Constrained Tracking(3)Rank-Constrained Tracking(4)Experiments – Non-rigid Body FactorizationSlide 29Slide 30Slide 31Slide 32Slide 33ConclusionThank you!!!Structure from MotionMani ThomasCISC 489/689RoadmapStructure from MotionBackgroundFeature based motion estimationKLT trackerFactorizationRigid body FactorizationMulti-body factorizationNon rigid body factorizationConclusionsStructure and Motion spaceStructure Estimation - BackgroundMany methods have been developed to tackle the Structure from Motion problemSzeliski and Kang, ‘93Azarbayejani and Pentland, ‘95Calway, ‘05Broida and Chellappa, ‘91Most methods involved an iterative minimization of an energy functionalComputationally expensiveFactorization – non-iterative solution to the structure from motion problemLeast squares estimate of structureComputationally inexpensivePriniciple of FactorizationFactorization principleStructure of an object resides in a low rank subspacePrincipal Component Analysis of the shape spaceCapture the subspace of the shape spaceReconstruct the structure using the projected components from this subspace instead of the actual spaceRank 3 subspace for a single rigid body (Tomasi and Kanade, ’92)Rank 6 subspace for scene with objects under constant velocity (Han and Kanade, ‘01)Rank 3K subspace for a non rigid body (Bregler et al., ’00)RoadmapStructure from MotionBackgroundFeature based motion estimationKLT trackerFactorizationRigid body FactorizationMulti-body factorizationNon rigid body factorizationConclusionsKanade-Lucas-Tomasi (KLT) trackerFeature based motion estimationExtract cornersCompute the motion parameters from the best bipartite graph Correspondence between the feature points in one image with those in the otherCorner extractionZ must be a well-conditioned 2 £ 2 matrixBoth the eigenvalues must be large and should not differ by several orders of magnitudeTwo small eigenvalues means a roughly constant intensity profileA large and a small eigenvalue corresponds to a unidirectional texture patternIn practice, when the smaller eigenvalue is large enough, Z is well conditionedmin(1, 2) > where is a predefined constantFor more information: J. Shi and C. Tomasi, “Good Features to Track”, CVPR ’94yyxyyxxxIIIIIIIIZKanade-Lucas-Tomasi (KLT) trackerKLT is a C implementation of a feature tracker for the computer vision communityPublic domain code that is maintained by S. Birchfield at http://www.ces.clemson.edu/~stb/klt/Available for Unix/Linux and Visual StudioDemo of the KLT program to track features across a sequence of imagesFor more information: J. Shi and C. Tomasi, “Good Features to Track”, CVPR ’94RoadmapStructure from MotionBackgroundFeature based motion estimationKLT trackerFactorizationRigid body FactorizationMulti-body factorizationNon rigid body factorizationConclusionsAffine CameraPix and Piy are the projection vectors that take the point [X, Y, Z] to (x, y)Pix and Piy are 1 £ 4 row vector, i is the frame number and j is one of the 3D pointAffine projection preserves the center of gravity 1 10001 1jjjyixiijijjjjyixiijijZYXvuZYXvuPPPP ~~44jjjyixiijijyiijxiijZYXvuPvPuPPAffine CameraRepeating the same with all the 3D pointsLHS is the collection of the points detected in the imageRHS has the 3D points with the projection matrix ~~~~~~P32121213222121PPPyixiPiPiiiPiiZZZYYYXXXvvvuuuPPRigid Body FactorizationDeveloped by Tomasi and Kanade in 1992Compute the measurement matrix WP feature points tracked across F framesPerform PCA using Singular Value Decomposition of the registered measurement matrixReconstruct the M (Motion) and S (Shape) using the top 3 singular valuesTPFPF 33332332ˆˆˆ~VUSMWTVUTWW ~PFPFFPFFFPFFPPvvvuuuvvvuuu332221211121111211SMW“Shape and Motion from Image Streams under Orthography: A Factorization Method”, IJCV, 1992Rigid Body Factorization (contd.)Rotation constraintsValid up to a linear transformation ARotational constraints to compute the parameters of AAAT is symmetricCan be estimated using eigen decompositionComputing the M and S after estimating the linear transformation APFPF 31333332332ˆˆSAAMSM011 fTTffTTffTTfjAAijAAjiAAiTPPFF 3213313333321333232ˆˆ,ˆˆVASAUMExperiments – Rigid Body FactorizationImplementation of the rigid body factorization30 Frames tracked using KLT trackerOnly tracks available over the entire sequence consideredMotion and Shape estimated using Singular Value DecompositionStructure estimated after computing the linear transformation Adata obtained from CMU image database http://vasc.ri.cmu.edu/idbExperiments – Rigid Body Factorization50 Frames tracked using KLT trackerEstimation of motion and structure using SVD of the tracked featuresViewing angle changed continuously to help visualize the hotel structuredata obtained from CMU image database http://vasc.ri.cmu.edu/idbRigid Body FactorizationAssumptions of the methodModeling of a Single BodyThe body is rigidThe camera projection is OrthographicRelaxation of the assumptionsSturm and Triggs, ’96 -
View Full Document