SURF-Face: Face Recognition Under Viewpoint
Consistency Constraints
Philippe Dreuw, Pascal Steingrube, Harald Hanselmann and Hermann Ney
Human Language Technology and Pattern Recognition, RWTH Aachen University, Aachen, Germany
Introduction Databases
I Most face recognition approaches are sensitive to registration errors I AR-Face
. rely on a very good initial alignment and illumination . variations in illumination
I We propose/analyze: . many different facial expressions
. grid-based and dense extraction of local features I CMU-PIE
. block-based matching accounting for different . variations in illumination (frontal images
viewpoints and registration errors from the illumination subset)
Feature Extraction Results: Manually Aligned Faces
Orig. IP Grid I AR-Face: 110 classes, 770 train, 770 test
I Interest point based feature extraction Descriptor Extraction # Features Error Rates [%]
. SIFT or SURF interest point detector Maximum Grid Grid-Best
. leads to a very sparse description SURF-64 IPs 164 × 5.6 (avg.) 80.64 84.15 84.15
SIFT IPs 128 × 633.78 (avg.) 1.03 95.84 95.84
I Grid-based feature extraction
SURF-64 64x64-2 grid 164 × 1024 0.90 0.51 0.90
. overlaid regular grid SURF-128 64x64-2 grid 128 × 1024 0.90 0.51 0.38
. leads to a dense description SIFT 64x64-2 grid 128 × 1024 11.03 0.90 0.64
U-SURF-64 64x64-2 grid 164 × 1024 0.90 1.03 0.64
U-SURF-128 64x64-2 grid 128 × 1024 1.55 1.29 1.03
U-SIFT 64x64-2 grid 128 × 1024 0.25 0.25 0.25
Feature Description
I CMU-PIE: 68 classes, 68 train (“one-shot” training), 1360 test
I Scale Invariant Feature Transform (SIFT)
Descriptor Extraction # Features Error Rates [%]
. 128-dimensional descriptor, histogram of gradients, scale invariant
Maximum Grid Grid-Best
I Speeded Up Robust Features (SURF)
SURF-64 IPs 164 × 6.80 (avg.) 93.95 95.21 95.21
. 64-dimensional descriptor, histogram of gradients, scale invariant SIFT IPs 128 × 723.17 (avg.) 43.47 99.33 99.33
I face recognition: invariance w.r.t. rotation is often not necessary SURF-64 64x64-2 grid 164 × 1024 13.41 4.12 7.82
. rotation dependent upright-versions U-SIFT, U-SURF-64, U-SURF-128 SURF-128 64x64-2 grid 128 × 1024 12.45 3.68 3.24
SIFT 64x64-2 grid 128 × 1024 27.92 7.00 9.80
U-SURF-64 64x64-2 grid 164 × 1024 3.83 0.51 0.66
Feature Matching U-SURF-128 64x64-2 grid 128 × 1024 5.67 0.95 0.88
U-SIFT 64x64-2 grid 128 × 1024 16.28 1.40 6.41
I Recognition by Matching
. nearest neighbor matching strategy
Results: Unaligned Faces
. descriptor vectors extracted at keypoints in a test image X are compared
to all descriptor vectors extracted at keypoints from the reference images I Automatically aligned by Viola & Jones I Manually aligned faces
Yn, n = 1, · · · , N by the Euclidean distance Descriptor Error Rates [%]
. decision rule: n AR-Face CMU-PIE
X o
X → r(X) = arg max max δ(xi, Yn,c) SURF-64 5.97 15.32
c n
xi∈X SURF-128 5.71 11.42 I Unaligned faces
SIFT 5.45 8.32
. additionally, a ratio constraint is applied in δ(xi, Yn,c)
U-SURF-64 5.32 5.52
I Viewpoint Matching Constraints U-SURF-128 5.71 4.86
. maximum matching: unconstrained U-SIFT 4.15 8.99
. grid-based matching: absolute box constraints
. grid-based best matching: absolute box constraints, overlapping
I Postprocessing Results: Partially Occluded Faces
. RANSAC-based outlier removal
I AR-Face: 110 classes, 110 train (“one-shot” training), 550 test
. RANSAC-based system combination
Descriptor Error Rates [%]
AR1scarf AR1sun ARneutral AR2scarf AR2sun Avg.
Matching Examples for the AR-Face and CMU-PIE Database SURF-64 2.72 30.00 0.00 4.54 47.27 16.90
SURF-128 1.81 23.63 0.00 3.63 40.90 13.99
Feature Maximum Grid Grid-Best Maximum Grid Grid-Best Feature SIFT 1.81 24.54 0.00 2.72 44.54 14.72
U-SURF-64 4.54 23.63 0.00 4.54 47.27 15.99
U-SURF-128 1.81 20.00 0.00 3.63 41.81 13.45
U-SIFT 1.81 20.90 0.00 1.81 38.18 12.54
U-SURF-128+R 1.81 19.09 0.00 3.63 43.63 13.63
U-SIFT+R 2.72 14.54 0.00 0.90 35.45 10.72
U-SURF-128+U-SIFT+R 0.90 16.36 0.00 2.72 32.72 10.54
SIFT SURF
Conclusions
I Grid-based local feature extraction instead of interest points
I Local descriptors:
. upright descriptor versions achieved better results
U-SIFT U-SURF . SURF-128 better than SURF-64
I Matching results for the AR-Face (left) and the CMU-PIE database (right) I System robustness: manually aligned/unaligned/partially occluded faces
. maximum matching show false classification examples . SURF more robust to illumination
. grid matchings show correct classification examples . SIFT more robust to changes in viewing conditions
. upright descriptor versions reduce the number of false matches I RANSAC-based system combination and outlier removal
Created with LATEXbeamerposter http://www-i6.informatik.rwth-aachen.de/~dreuw/latexbeamerposter.php
http://www-i6.informatik.rwth-aachen.de <surname>@cs.rwth-aachen.de
