440 likes | 451 Views
Faces in the Wild. Detection, Alignment and Recognition of Real World Faces Erik Learned-Miller with Vidit Jain, Gary Huang, Andras Ferencz, et al. Is Face Recognition Solved?. Is Face Recognition Solved?. “100% Accuracy in Automatic Face Recognition” [!!!]. Science 25 January 2008.
E N D
Faces in the Wild Detection, Alignment and Recognition of Real World Faces Erik Learned-Miller with Vidit Jain, Gary Huang, Andras Ferencz, et al.
Is Face Recognition Solved? “100% Accuracy in Automatic Face Recognition” [!!!] Science 25 January 2008
Is Face Recognition Solved? “100% Accuracy in Automatic Face Recognition” [!!!] Science 25 January 2008 A history of overstated results.
The Truth • Many different face recognition problems • Out of context, accuracy is meaningless! • Many problems are REALLY HARD! • For some problems state of the art is 70% or worse! • We have a long way to go!
Face Recognition at UMass • Labeled Faces in the Wild • The Detection-Alignment-Recognition pipeline • Congealing and automatic face alignment • Hyper-features for face recognition • New directions in recognition
Labeled Faces in the Wild http://vis-www.cs.umass.edu/lfw/
The Many Faces of Face Recognition Labeled Faces in the Wild
The Many Faces of Face Recognition Labeled Faces in the Wild
The Many Faces of Face Recognition Labeled Faces in the Wild
The Many Faces of Face Recognition Labeled Faces in the Wild
The Many Faces of Face Recognition Labeled Faces in the Wild
Labeled Faces in the Wild • 13,233 images, with name of each person • 5749 people • 1680 people with 2 or more images • Designed for the “unseen pair matching problem”. • Train on matched or mismatched pairs. • Test on never-before-seen pairs. • Distinct from problems with “galleries” or training data for each target image. • Best accuracy: currently about 73%!
Detection-Alignment-Recognition Pipeline Detection Alignment Recognition “Same”
Detection-Alignment-Recognition Pipeline Detection Alignment Recognition “Same” Parts should work together.
Labeled Faces in the Wild • All images are output of a standard face detector. • Also provides aligned images. • Consequence: any face recognition algorithm that works well on LFW can easily be turned into a complete system.
Criterion of Joint Alignment • Minimize sum of pixel stack entropies by transforming each image. A pixel stack
Congealing Complex Images Window around pixel SIFT vector and clusters SIFT clusters vector representing probability of each cluster, or “mixture” of clusters
Martian training set = Test: Find Bob after one meeting ? = = Crash Course on Martian Identification Bob
Training Data “same” “different”
General Approach to Hyper-feature method • Carefully align objects • Develop a patch-based model of image differences. • Score match/mismatch based on patch differences.
Three Models • Universal patch model: P(patchDistance|same) P(patchDistance|different) • Spatially dependent patch model: P(patchDistance |same,x,y) P(patchDistance |different,x,y) • Hyper-feature dependent model: • P(patchDistance |same,x,y,appearance) • P(patchDistance |different,x,y,appearance)
Universal Patch Model A single P(dist | same) for all patches Different blue patches are evidence against a match!
Spatial Patch Model P(dist|same,x1,y1) estimated separately from P(dist|same,x2,y2) Greatly increases discriminativeness of model.
Hyper-Feature Patch Model Is the patch from a matching face going tomatch this patch?
Hyper-Feature Patch Model Is the patch from a matching face going tomatch this patch? Probably yes
Hyper-Feature Patch Model What about this patch?
Hyper-Feature Patch Model What about this patch? Probably not.
Ridiculous Errors from the World’s Best Unconstrained Face Recognition System
Ridiculous Errors from the World’s Best Unconstrained Face Recognition System
The New Mission: Estimate Higher Level Features Can we guess pose?
The New Mission: Estimate Higher Level Features Can we guess gender?
The New Mission: Estimate Higher Level Features Can we guess degree of balding, beardedness, moustache?
The New Mission: Estimate Higher Level Features Can we say that none of these individuals are the same person?
Who’s This? from www.coolopticalillusions.com