[Computer Vision 7] Adaptive As-Natural-As-Possible Image Stitching

Local Homography Model

Only in overlapping region: The local homograph is estimated by including a offset weight.

H at the location p_j: h_j = argmin Σ ω_i,j ||Ah_j||^2 (p_i: matching points, Wj =  diag([ω_i,j]):  high value for the neighbors of p_j, equal values for those are very far)

Homography Linearization

In non-overlapping region: To reduce the distortions, linearize non-overlapping region using Taylor series of homographic transformation H. (set anchor points as the boundary of overlapping region, and compute a weighted average of the transformations)

Global Similarity Transformation

To further reduce the distortions, use a similarity transformation in the non-overlapping areas. 

Determine global similarity -->choose the one with smallest rotation angle