Optimal Features ASM Texture description based on Taylor series Grids centered at the landmarks for local analysis Non linear classifier (kNN) for inside-outside labeling inside outside B. van Ginneken, A.F. Frangi, J.J. Staal, B.M. ter Haar Romeny, and M.A. Viergever (2002) IEEE Transactions on Medical Imaging, 21(8):924–933
Optimal Features ASM Face is too complex for the proposed labeling Thin zones generate profile variations Classes unbalance in high curvature points kNN slow (set dependent) Image features dependent on rotation 1 2 2
Invariant Optimal Features ASM F.M. Sukno, S. Ordas, C. Butakoff, S. Cruz, and A.F. Frangi IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(7):1105–1117
Invariant Optimal Features ASM Distance-based labeling 180 profiles of man and women with IOF-ASM 1 2 4
Invariant Optimal Features ASM 1 2 Multi-valued neuron classifier Single neuron Very fast Appropriate combination of derivatives allows for invariance to rigid transformations i 0 1 k-1 k- 2 Z 5
Segmentation tests Experiments on images Point to curve error Point to point error
IOFASM vs ASM DatasetImagesError AR % Equinox % XM2VTS % 1 2 7
IOFASM vs ASM ASM IOF-ASM 1 2 8
Identity Verification: Texture Based on texture Eigenfaces-like approach from the segmentation results 1 2 9
Identity Verification: Texture
Related work
Conclusions on IOF-ASM 1 2 By using more elaborate descriptions of the texture it is possible to increase the accuracy of ASMs IOF-ASM provides a generic framework Features are optimized for every landmark Allows for a trade off between accuracy and speed Feature selection: –15% error / –50% time About 30% more accurate than ASM in facial feature localization Derives in better identification rates Invariant to in-plane rotations 12
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