Guillaume-Alexandre Bilodeau

Slides:



Advertisements
Similar presentations
Pseudo-Relevance Feedback For Multimedia Retrieval By Rong Yan, Alexander G. and Rong Jin Mwangi S. Kariuki
Advertisements

Université du Québec École de technologie supérieure Face Recognition in Video Using What- and-Where Fusion Neural Network Mamoudou Barry and Eric Granger.
Recognizing Human Actions by Attributes CVPR2011 Jingen Liu, Benjamin Kuipers, Silvio Savarese Dept. of Electrical Engineering and Computer Science University.
Loris Bazzani*, Marco Cristani*†, Alessandro Perina*, Michela Farenzena*, Vittorio Murino*† *Computer Science Department, University of Verona, Italy †Istituto.
Zhimin CaoThe Chinese University of Hong Kong Qi YinITCS, Tsinghua University Xiaoou TangShenzhen Institutes of Advanced Technology Chinese Academy of.
Principal Component Analysis Based on L1-Norm Maximization Nojun Kwak IEEE Transactions on Pattern Analysis and Machine Intelligence, 2008.
Face Recognition. Introduction Why we are interested in face recognition? Why we are interested in face recognition? Passport control at terminals in.
Proposed concepts illustrated well on sets of face images extracted from video: Face texture and surface are smooth, constraining them to a manifold Recognition.
Patch to the Future: Unsupervised Visual Prediction
Intelligent Systems Lab. Recognizing Human actions from Still Images with Latent Poses Authors: Weilong Yang, Yang Wang, and Greg Mori Simon Fraser University,
Adviser:Ming-Yuan Shieh Student:shun-te chuang SN:M
Robust Object Tracking via Sparsity-based Collaborative Model
Multi-View Learning in the Presence of View Disagreement C. Mario Christoudias, Raquel Urtasun, Trevor Darrell UC Berkeley EECS & ICSI MIT CSAIL.
Watching Unlabeled Video Helps Learn New Human Actions from Very Few Labeled Snapshots Chao-Yeh Chen and Kristen Grauman University of Texas at Austin.
São Paulo Advanced School of Computing (SP-ASC’10). São Paulo, Brazil, July 12-17, 2010 Looking at People Using Partial Least Squares William Robson Schwartz.
Real-time Embedded Face Recognition for Smart Home Fei Zuo, Student Member, IEEE, Peter H. N. de With, Senior Member, IEEE.
1 Learning to Detect Objects in Images via a Sparse, Part-Based Representation S. Agarwal, A. Awan and D. Roth IEEE Transactions on Pattern Analysis and.
Generic Object Recognition -- by Yatharth Saraf A Project on.
A Study of Approaches for Object Recognition
UPM, Faculty of Computer Science & IT, A robust automated attendance system using face recognition techniques PhD proposal; May 2009 Gawed Nagi.
Jacinto C. Nascimento, Member, IEEE, and Jorge S. Marques
Oral Defense by Sunny Tang 15 Aug 2003
REALTIME OBJECT-OF-INTEREST TRACKING BY LEARNING COMPOSITE PATCH-BASED TEMPLATES Yuanlu Xu, Hongfei Zhou, Qing Wang*, Liang Lin Sun Yat-sen University,
Partial Face Recognition
Wang, Z., et al. Presented by: Kayla Henneman October 27, 2014 WHO IS HERE: LOCATION AWARE FACE RECOGNITION.
Thien Anh Dinh1, Tomi Silander1, Bolan Su1, Tianxia Gong
Mining Discriminative Components With Low-Rank and Sparsity Constraints for Face Recognition Qiang Zhang, Baoxin Li Computer Science and Engineering Arizona.
Kourosh MESHGI Shin-ichi MAEDA Shigeyuki OBA Shin ISHII 18 MAR 2014 Integrated System Biology Lab (Ishii Lab) Graduate School of Informatics Kyoto University.
Marcin Marszałek, Ivan Laptev, Cordelia Schmid Computer Vision and Pattern Recognition, CVPR Actions in Context.
Window-based models for generic object detection Mei-Chen Yeh 04/24/2012.
Pairwise Linear Regression: An Efficient and Fast Multi-view Facial Expression Recognition By: Anusha Reddy Tokala.
1 Webcam Mouse Using Face and Eye Tracking in Various Illumination Environments Yuan-Pin Lin et al. Proceedings of the 2005 IEEE Y.S. Lee.
Face Recognition: An Introduction
Face Detection Ying Wu Electrical and Computer Engineering Northwestern University, Evanston, IL
Geodesic Flow Kernel for Unsupervised Domain Adaptation Boqing Gong University of Southern California Joint work with Yuan Shi, Fei Sha, and Kristen Grauman.
HIGH PERFORMANCE OBJECT DETECTION BY COLLABORATIVE LEARNING OF JOINT RANKING OF GRANULES FEATURES Chang Huang and Ram Nevatia University of Southern California,
GENDER AND AGE RECOGNITION FOR VIDEO ANALYTICS SOLUTION PRESENTED BY: SUBHASH REDDY JOLAPURAM.
Irfan Ullah Department of Information and Communication Engineering Myongji university, Yongin, South Korea Copyright © solarlits.com.
Intelligent Database Systems Lab 國立雲林科技大學 National Yunlin University of Science and Technology 1 Recognizing Partially Occluded, Expression Variant Faces.
Max-Confidence Boosting With Uncertainty for Visual tracking WEN GUO, LIANGLIANG CAO, TONY X. HAN, SHUICHENG YAN AND CHANGSHENG XU IEEE TRANSACTIONS ON.
Face Detection 蔡宇軒.
1 Bilinear Classifiers for Visual Recognition Computational Vision Lab. University of California Irvine To be presented in NIPS 2009 Hamed Pirsiavash Deva.
Face recognition using Histograms of Oriented Gradients
Abdul Jabbar Siddiqui, Abdelhamid Mammeri, and Azzedine Boukerche
Face Detection EE368 Final Project Group 14 Ping Hsin Lee
Bag-of-Visual-Words Based Feature Extraction
A Pool of Deep Models for Event Recognition
Learning Mid-Level Features For Recognition
Article Review Todd Hricik.
PRESENTED BY Yang Jiao Timo Ahonen, Matti Pietikainen
Face Recognition and Feature Subspaces
Fast Preprocessing for Robust Face Sketch Synthesis
Hybrid Features based Gender Classification
Recovery from Occlusion in Deep Feature Space for Face Recognition
Unsupervised Face Alignment by Robust Nonrigid Mapping
Cheng-Ming Huang, Wen-Hung Liao Department of Computer Science
Facial Recognition in Biometrics
Hu Li Moments for Low Resolution Thermal Face Recognition
Domingo Mery Department of Computer Science
Liyuan Li, Jerry Kah Eng Hoe, Xinguo Yu, Li Dong, and Xinqi Chu
George Bebis and Wenjing Li Computer Vision Laboratory
Housam Babiker, Randy Goebel and Irene Cheng
Speaker: YI-JIA HUANG Date: 2011/12/08 Authors: C. N
A Novel Smoke Detection Method Using Support Vector Machine
Recognizing Deformable Shapes
Domingo Mery Department of Computer Science
Human-object interaction
Volume 50, Issue 1, Pages (April 2006)
Presentation transcript:

Guillaume-Alexandre Bilodeau Dynamic Selection of Exemplar-SVMs for Watch-List Screening Through Domain Adaptation Saman Bashbaghi Eric Granger Robert Sabourin Guillaume-Alexandre Bilodeau École de technologie supérieure, Université du Québec December 21, 2016

Outline Introduction Application, Challenges, Objectives, and Contributions Ensembles of Exemplar-SVMs Through Domain Adaptation Design phase (pool generation) Operational phase (dynamic selection and integration) Experimental Results COX-S2V data set Experimental protocol Simulation results Conclusions

1. Introduction Application Area Generic system for video-based face recognition

1. Introduction Application Area Video Surveillance applications detect the presence of target individuals of interest (enrolled to the system) based on their facial model types of systems (based on facial model): video-to-video FR (e.g., face re-identification, search and retrieval) still-to-video FR (e.g., watch-list screening) individual-specific detection among a restrained list of individuals [Pagano et al., 2012] Watch-List Screening recognize target individuals from watch-list images, under semi- and unconstrained environments [Bashbaghi et al., 2014] face model is defined as either a set of reference samples (stored in a gallery for template matching), or a set of classifier parameters estimated from reference samples (for pattern classification)

1. Introduction Challenges Limited and Imbalanced Data design with limited number of target reference face captures and an abundance of non-target references small number of target inviduals appearing during operations can estimate of parameters (thresholds, feature subsets, classifiers) with auxiliary data obtained from videos of non- target people in the scene Face Capture and Classification different cameras used to capture reference faces (still camera under controlled conditions) and operational faces (IP video camera under semi- or uncontrolled conditions) nuisance factors – variation in illumination, pose, scale, expression, occlusion, contrast, focus, sharpness, etc.

1. Introduction Objectives Design robust still-to-video FR systems for watch- list screening approaches: provide a novel dynamic ensemble selection method for still-to-video FR through individual-specific multi- classifier system exploit classification systems through domain adaptation with the availability of a single labeled target ROI obtained from a high-quality still reference and unlabeled non-target ROIs captured either from the cohort or low-quality calibration videos evaluate the performance of different training schemes and dynamic selection of ensembles

1. Introduction Contributions A robust still-to-video FR can be designed under SSPP using: Multiple and diverse face representations uniform patches Random Subspace Method provide ensemble diversity and improve robustness to various perturbation factors frequently occur in video surveillance environments Ensemble of e-SVMs through domain adaptation discriminate between a single labeled high-quality target ROI and abundant of unlabeled non-target ROIs from calibration videos Dynamic Classifier Selection An intuitive approach to provide the best separation w.r.t. non-target samples in the feature space that allows the system to select classifiers properly fit the acquisition conditions

2. Dynamic Ensemble of SVM Classifiers Operational system Block diagram of the proposed still-to-video FR system using dynamic ensemble of e-SVMs per target individual

2. Dynamic Ensemble of SVM Classifiers Design Phase (Generating a pool) Multiple Diverse Face Representations Uniform Patches local parts of face to provide diversity for ensembles are suitable for partially occluded faces Feature Extraction Technique HOG, LPQ Random Subspace Method allow for randomly sampling and selection different feature subsets from feature space of input samples Classification Systems Exemplar-SVM relies on a single labeled still face from the enrollment domain along with unlabeled faces in calibration videos from the operational domain

2. Dynamic Ensemble of SVM Classifiers Enrollment Phase (Training schemes) (a) Scheme 1 (b) Scheme 2 (c) Scheme 3 Illustration of e-SVMs in a 2D feature space trained using different classification schemes according. (a) a labeled target still vs labeled non-target still ROIs of Ed, (b) a labeled target still vs unlabeled non-target video ROIs of Od, and (c) a labeled target still vs camera-specific unlabeled non-target video ROIs.

2. Dynamic Ensemble of SVM Classifiers Operational Phase (Dynamic classifier selection) Illustration of the proposed dynamic classifier selection approach in a 2D feature space

3. Experimental Results Dataset COX-S2V dataset [Huang et al., 2013] 1000 subjects, where each subject has a high-quality still image, and four lower-quality facial trajectories (each trajectory has 25 faces of 16x20 and 48x60 resolutions) Example of the reference still of subject ID \#1 and some corresponding videos from the COX-S2V dataset

3. Experimental Results Experimental methodology Protocol for independent 5 trials Design phase (randomly select) 20 individuals in the watch-list (20 high-quality stills) 100 unknown person to form UBM (ROIs video capture) Operational phase (test mode) 100 other unknown persons (ROIs video capture) ROIs video capture of individuals of interest (one at a time) Measures: transaction-based analysis (matching performance) pAUC(20%), AUPROC Time Complexity: Number of dot products required by the ensemble of e-SVMs to process a probe

3. Experimental Results Transaction-level Average transaction-level performance and complexity of the proposed system against state-of-the-art s Average pAUC(20\%) and AUPR performance of different training schemes at transaction-level Average performance of the proposed system with or without dynamic selection

4. Conclusion Training schemes 1 is greatly outperformed by schemes 2 and 3, where calibration videos from operational domains are employed for domain adaptation to train e-SVMs In the scheme 2, videos from all of the cameras are employed to generate an e-SVM pool, while 4 camera- specific e-SVM pools are generated for scheme 3 using videos of each camera Dynamic selection can improve the performance instead of combining all of classifiers, where it is efficient and does not significantly increase the computational burden

Thanks for your attention