Université du Québec École de technologie supérieure Face Recognition in Video Face Recognition in Video Using What-and-Where Fusion Using What-and-Where Fusion Neural Network Neural Network Mamoudou Barry and Eric Granger Mamoudou Barry and Eric Granger Laboratoire d’imagerie, de vision et Laboratoire d’imagerie, de vision et d’intelligence artificielle d’intelligence artificielle École de technologie supérieure École de technologie supérieure Montreal, Canada Montreal, Canada
22
Embed
Université du Québec École de technologie supérieure Face Recognition in Video Using What- and-Where Fusion Neural Network Mamoudou Barry and Eric Granger.
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Université du Québec
École de technologie supérieure
Face Recognition in Video Using What-Face Recognition in Video Using What-and-Where Fusion Neural Networkand-Where Fusion Neural Network
Mamoudou Barry and Eric GrangerMamoudou Barry and Eric GrangerLaboratoire d’imagerie, de vision et d’intelligence Laboratoire d’imagerie, de vision et d’intelligence
artificielleartificielleÉcole de technologie supérieureÉcole de technologie supérieure
Challenges of video-based face recognitionChallenges of video-based face recognition
low quality and resolution of frames.low quality and resolution of frames.
uncontrolled environments: variation in poses, uncontrolled environments: variation in poses, orientation, expressions, illumination, occlusion, orientation, expressions, illumination, occlusion, etc.etc.
Université du Québec
École de technologie supérieure4
1. Introduction1. Introduction
General system for face recognition in videoGeneral system for face recognition in video
Université du Québec
École de technologie supérieure5
1. Introduction1. Introduction
State of the artState of the art
1.1. Methods based on static imagesMethods based on static images– exploit quality metric, and recognize only high exploit quality metric, and recognize only high
quality ROIsquality ROIs
2.2. Spatiotemporal approachesSpatiotemporal approaches– track faces in the environment, and recognize track faces in the environment, and recognize
individuals over several samplesindividuals over several samples
Université du Québec
École de technologie supérieure6
1. Introduction1. Introduction
ObjectivesObjectives
Observe the effectiveness of the What-and-Where Observe the effectiveness of the What-and-Where fusion neural network in video-based face recognitionfusion neural network in video-based face recognition
Robust operation in uncontrolled environmentsRobust operation in uncontrolled environments
1.1. WhatWhat data data:: intrinsic intrinsic properties of a properties of a face face (to classifier)(to classifier)
2.2. WhereWhere data data:: ccontextual ontextual information information (to tracker)(to tracker)
Tracker
Classifier
1
h
R
1
k
L
1
k
L
1
k
L
1
k
L
Evidenceaccumulation
track#
WHAT data stream
WHERE data stream
yeyab
Fe1
Feh
FeR
Université du Québec
École de technologie supérieure8
TrackerTracker: : bank of Kalman filtersbank of Kalman filters estimates the future position stimates the future position
of faces in a sceneof faces in a scene
ClassifierClassifier: : fuzzy ARTMAPfuzzy ARTMAP classifies faces detected in a sceneclassifies faces detected in a scene neural network architecture capable neural network architecture capable of of
fast, stable, online, unsupervised fast, stable, online, unsupervised or or supervised, incremental learning, supervised, incremental learning, classification classification and predictionand prediction
train:train: train fuzzy ARTMAP with train fuzzy ARTMAP with What What data, data, using two training strategies using two training strategies Hold-Out Validation (HV)Hold-Out Validation (HV) Particle Swarm Optimization (PSO) to optimize hyper-Particle Swarm Optimization (PSO) to optimize hyper-
parameters (Granger parameters (Granger et al.,et al., 2007) 2007)
testtest: : classify classify What What data with fuzzy ARTMAP and data with fuzzy ARTMAP and track track Where Where data with Kalman filtersdata with Kalman filters
accuracy:accuracy: average classification error (estimate of average classification error (estimate of generalization error)generalization error)
resource requirements:resource requirements:
compression: compression: average number of training patterns average number of training patterns per categoryper category
convergence time:convergence time: average number of epochs average number of epochs required to complete learning.required to complete learning.
Université du Québec
École de technologie supérieure15
4. Results4. Results
Examples of Face DetectionsExamples of Face Detections
Université du Québec
École de technologie supérieure16
4. Results4. Results
Average error and compressionAverage error and compressionvsvs. . ROI scaling size (with 100% of training data)ROI scaling size (with 100% of training data)
Université du Québec
École de technologie supérieure17
4. Results4. Results
Average error and compressionAverage error and compressionvsvs. training subset size (with a |ROI| =10x10). training subset size (with a |ROI| =10x10)
Université du Québec
École de technologie supérieure18
4. Results4. Results
Average convergence timeAverage convergence time
fuzzy ARTMAP with HV: ~fuzzy ARTMAP with HV: ~1 epoch1 epoch
fuzzy ARTMAP with PSO: ~fuzzy ARTMAP with PSO: ~543 epochs543 epochs
(60 particles x ~8.9 iterations x 1 epoch)(60 particles x ~8.9 iterations x 1 epoch)
Example of prediction errors over timeExample of prediction errors over time
4. Results4. Results
Université du Québec
École de technologie supérieure21
Effectiveness of the What-and-Where fusion neural Effectiveness of the What-and-Where fusion neural network in improving the accuracy on complex video data network in improving the accuracy on complex video data (about 50% over fuzzy ARTMAP alone, and k-NN).(about 50% over fuzzy ARTMAP alone, and k-NN).
The system is less sensitive to noise: attenuation of fuzzy The system is less sensitive to noise: attenuation of fuzzy ARTMAP poor predictions.ARTMAP poor predictions.
Optimizing the network internal parameters using PSO Optimizing the network internal parameters using PSO learning strategy improves the accuracy of the system.learning strategy improves the accuracy of the system.
Fuzzy ARTMAP yields a higher compression than k-NN: Fuzzy ARTMAP yields a higher compression than k-NN: suitable for real time and ressource limited applications.suitable for real time and ressource limited applications.
5. Conclusion5. Conclusion
Université du Québec
École de technologie supérieure22
6. Future work6. Future work
Explore different ARTMAP models to Explore different ARTMAP models to improve the classification rate.improve the classification rate.
Explore other representations (features) of face Explore other representations (features) of face based on biological vision perception.based on biological vision perception.
Investigate for more robust tracking algorithms Investigate for more robust tracking algorithms such as Extended Kalman filter, Particle filters, such as Extended Kalman filter, Particle filters, etc., for non linear tracking.etc., for non linear tracking.