1. Facial Expression
Recognition in the Wild
Adil Khan
Machine Learning & Knowledge Representation Lab
Institute of AI & Robotics
Innopolis University
2. About ME!
Associate Prof. Innopolis University (Russia)
Head of Machine Learning & Knowledge
Representation Lab
Machine Learning, Computer Vision, Pervasive Computing
3. Why FER?
We are emotional beings! Happy → Pizza BUT Sad → Ice-cream
4. Why FER?
Most Effective Way to Understand Emotions:
A. Mehrabian, Nonverbal Communication. New Brunswick, NJ, USA: Aldine, 2007.
7. FER: An Extensively Researched Area!
Facial expression recognition using
radial encoding of local Gabor features
and classifier synthesis
Compound facial expressions of
emotion
Rapid perceptual integration of facial
expression and emotional body
language
Simulationist models of face-based
emotion recognition
Emotion recognition in human
computer interaction
Real-time Mobile Facial Expression
Recognition System
13. FER Pipeline
Pre-Processing Face Detection Facial Components
Feature Extraction Classification Recognized
Expression Class
Image or
Video
14. Journey Begins With Data!
https://link.springer.com/article/10.1007/s11042-016-4321-2
15. We Combine but Combine Intelligently!
http://www.araya.org/wp/wp-content/uploads/2016/10/gan.png
16. Features: Are Very Important; Already Got Some!
Two broad categories:
1. Geometric Features
ASM, AAM
2. Appearance based features
GF, LBP, LDA, HOGs, etc.
17. Feature Fusion Has Been Tried; Works to Some
Extent
1. Simple concatenation
2. Spectral Embedding
3. Genetic Algorithms
4. Multi Kernel Learning
5. Ensemble of classifiers
6. etc.
Fusion
Shape Info Texture Info
Fused Feature Vector
20. Two Level is Good, but Too Naive/Heuristic!
http://uclab.khu.ac.kr/resources/publication/J_201.pdf
21. We combine: One Network for Both Tasks!
A Deep Network
Cluster learning at the first layers
Representation Learning within each cluster at the subsequent layers
Classification at the last layer
But have to figure out how to backprop
22. We Test on the REAL Data!
The biggest challenge, but we are trying
We collect our own real, spontaneous expression data
1. Online resources
2. Offline resources