Out-of-plane Rotations Environment constraints ● Surveillance systems ● Car driver images ASM: ● Similarity does not remove 3D pose ● Multiple-view database Other approaches ● Non-linear models ● 3D models: multiple views Database
Projective Geometry Geometric operations by means of linear algebra 2D points are 3- component vectors Multiple views of the same planar object can be related by homographies
Homographies Homographies hold both for object or camera movements The points must be coplanar H
Coplanar face model Silhouette points are excluded (out of main plane) Half the nose points are excluded (easy occlusion) First iteration: At least 8 correspondences to compute H (4 2D-points) Model Coordinates Image Coordinates
Image Matching ASM Image Model (Similarity) Gradient normal to the shape contour Projective transformations Do not preserve angles nor distance relationships H
Database (40 people)
Results PASM ASM Training and test on multi-view data Cross validation 7
Comparison to related work Ratios with respect to error on frontal images 8
Results training just a single view (frontal) Training set: Frontal Test set: Multilple views 9
Analysis of the single-view case
Conclusions on PASM If multi-view dataset available Almost invariant to rotations up to 60 degrees Training only on frontal views Considerably reduces (50%) variation of ASM due to viewpoint Left-right rotations better handled than up-down nodding Very difficult to compare to other results Points used for alignment can affect performance Not considerable for expected ASM precision 11
How reliable is the result?