[Computer Vision 10] Shape and motion from image streams under orthography: A factorization method

The Rank Theorem

Registered Measurement Matrix D_2m*n: m views, n points (normalized points: coordinate origin locates at the centroid of 3D object); --> Singular value decomposition: D is at most of rank 3 (image measurements in D are highly redundant)

D = M*S: M_2m*3: camera rotation, S_3*n: shape in a coordinate system attached to the object centroid

The Metric Constraints

D = M*S: M and S are not unique. M = M_hat*Q, S = Q^-1*S_hat: M_hat and S_hat are the linear transformation of true motion and shape matrices.

Axes of camera frame (rows of M) is orthogonal and normalized.

i_hat^T * Q * Q^T * i_hat = 1

j_hat^T * Q * Q^T * j_hat = 1

i_hat^T * Q * Q^T * j_hat = 0

Occlusions

The factorization method only be applied to the sub-measurement matrix with completed values --> Obtain full motion matrix and partial shape matrix/ partial motion matrix and full shape matrix --> Reconstruction