MMM2020 Tutorial Session

26th International Conference on Multimedia Modeling (MMM 2020)

Jan 5 (Sun) Tutorial Session Contents Summary

Tutorial Ⅰ : Introduction to biometrics and anti-spoofing

1. Biometric Applications : Finger Print / Iris Recognition / Face / Vein
   • Finger Print : widely used and still valuable
     • 6 HR Baby -> F.P. Read
     • Mobile :
       – scanned only small part of F.P. -> it drops off accuracy
       – rotations
       – spoofing attack
     • Deep Learning ?
       – There is security reason why hand-crafted features are still used
     • Algorithms
       – Point-based matching algorithm
       > Minutiae points
       > enrollment + stitching
   • Iris Recognition
     • Main Problem : Glass Occlusion & Secular Reflection
     • One landmark : Daugman’s model (patent expired at 2011)
       1) Iris Segmentation
       2) Rubber-Sheet model
       3) Feature Extraction
   • Face Recognition
     • DeepFace Algorithm (Facebook 2014)
     • Almost mature
     • Accuracy on LFW DB > 99% (almost perfect)
   • Vein Recognition
     • Under the skin
     • start to focus on vein due to its anti-spoofing ability
     • Line Pattern vs. Texture Pattern
     • Wrist Vein -> Bio Watch
2. Anti-spoofing
   • FP
     • Local textual patterns are employed as features
   • Face
     • Photo, Video, Mask
     • Anti-spoofing
       – Motion based
       – Textural difference
       – Deep Neural Network based ( Binary Classification )
       ※ Very week for unseen dataset

Tutorial Ⅱ : Recent advances in deep novelty detection for medical imaging

1. Medical Imaging
   • CT, MRI, ultrasonic, …
   • Anomalies : Uncommon patterns in images
   • Anomaly Detection in Medical Imaging
     • Early Detection
     • Problems in real world
2. Anomaly Detection System (ADS )
   • Localize anomalies and quantify their differences
   • Three Approaches
     • Supervised
       – “Chest X-ray 14” Dataset
       > kaggle dataset [Link]
       > 112,120 images
     • Unsupervised
       – Abnormal vs Normal (clustering)
       – difficult to recognize without expertise
       – require to define dissimilarity measures
     • Semi-Supervised
       – modeling normality only
3. Challenges in Anomaly Detection
   • Modeling methods of normal cases
     • Representation learning of medical images
       – Curse of dimensionality : 3D/4D volume (CT, MRI), large 2D image (X-Ray)
     • Lack of well-defined representative boundary of normal cases
   • Data collection from healthy subjects
     • Good spread of normal cases : represent the diversity of target population
     • Factors to consider
       – Age, gender, imaging modality, hospital, and imaging protocol
   • Generalization of models
     • How to build more generalized models of normality with limited sample
     • Data augmentation, hyperparameter settings, adding noise, …
   • Anomaly scores
     • Margin from the boundaries of normal cases
     • Evaluation : Pixel-wise, patch-wise, region-wise, image-wise
4. Models
   • Auto Encoders
     • Robust Deep Autoencoders (robust PCA)
     • Deep Autoencoding Gaussian Mixture Model
     • Adversarial Autoencoder
   • GAN
     • Deep convolutional GAN (AnoGAN)
     • GANomaly
   • One-class classification Methods
     • Deep SVDD (Support Vector Data Description)
     • OC – CNN (one-class CNN)
     • Adversarially Learned One-class classifier for Novelty Detection

Tutorial Ⅲ : Haptic Interaction with multimedia data

1. Haptic Actuator, 양태헌, 고려대
   • Apple’s TAPTIC engine
     • Designed to big
   • Touch Pad
     • Taptic Engine + Force Sensor
       small horizontal actuation creates illusion of button sensation
   • Smart Fluid-based haptic
2. Electroactive Polymer Actuator, 김상윤, 경희대
   • Kinesthetic sensation
   • Tactile sensation
   • CES 2019 Trend : Rigid -> Flexible
   • Wave-form ePVC gel => to increase vibration
3. Haptic Rendering, 전석희, 경희대 ( 소개 링크: [Link] )
   • VR game player since 2013
   • Two purpose of haptic system
     • Reproduction of real ( in VR or Teleoperation )
     • Another information ( in HCI )
   • Three kinds of Haptics
     • Human Haptics
     • Computer Haptics
       – Haptic Rendering
       – Haptic Modeling
     • Machine Haptics
   • Haptic Rendering
     • Collision Detection
       – Much Faster than Computer Graphics << 1 ms
       – All computation should be done within 1 ms
   • Haptic Modeling
     • Stiffness, texture, geometric modeling, friction, inertia, weight