Report

VGG reading group presentation by Minh Hoai Cumulative Attribute Space for Age and Crowd Density Estimation Ke Chen1, Shaogang Gong1, Tao Xiang1, Chen Change Loy2 1. Queen Mary, University of London 2. The Chinese University of Hong Kong Tasks How old are they? How many people? What is the head angle? A Regression Framework Input images Features AAM feature Feature extraction Segment feature Edge feature Texture feature Label space Learn the mapping Regression Label (age, count) Challenge – Sparse and Unbalanced data Data distribution of FG-NET Dataset Challenge – Sparse and Unbalanced data Data distribution of UCSD Dataset Proposed Approach Solution: • Attribute Learning can address data sparsity problem - Exploits the shared characteristics between classes Has sematic meaning Question to address: • How to exploit cumulative dependent nature of labels in regression? …… …… Age 20 Age 21 …… Age 60 Cumulative Attribute Cumulative attribute (dependent) 0 1 … 20 1 … Age 20 1 0 Vs. 20th 1 0 0 0 0 0 … … the rest Non-cumulative attribute (independent) 0 Limitation of Non-cumulative Attribute 0 00 1 00 0 1 0 0 1 0 0 …… 0 … 0 … … … 0 … Age 20 … 20th 00 0 00 21st Age 60 21 60th Advantages of Cumulative Attribute 1 1 1 … … 20 1 1 0 the rest 0 … 0 1 0 1 0 0 …… 0 1 … … Age 20 40 attributes 1 attribute changes change 0 21 60 60 Age 21 Proposed Framework xi The task yi Label (e.g., age) Feature vector (e.g., intensity) 1 Regressor 1 1 2 … Cumulative attribute yi 0 … ai 1 0 Regressor Proposed Framework xi Our task How are these regressors learned? yi Label (e.g., age) Feature vector (e.g., intensity) 1 Regressor 1 1 2 … See next slide! Cumulative attribute yi 0 … ai 1 0 Regressor Can use any regression method: Support Vector Regression, Ridge Regression Regressor for Cumulative Attributes # of training data Regularization Closed-form solution: Cumulative attribute Image feature vector Regression error Parameters to learn Experiments Baseline Methods and Name Abbreviation CA-SVR Cumulative attributes 1 2 1 1 … yi 1 0 … 0 SVR xi yi Support Vector Regression (SVR) Feature vector Label SVR NCA-SVR Non-Cumulative attributes 1 2 0 0 … yi 1 0 … 0 Cumulative (CA) vs. Non-cumulative (NCA) Mean absolute error (lower is better) Percentage of prediction within 5 years (higher is better) Age Estimation Cumulative (CA) vs. Non-cumulative (NCA) Mean absolute error (lower is better) Mean squared error (lower is better) Crowd Counting Mean deviation error (lower is better) Crowd Counting Results Based on regression Ridge Regression without attributes Proposed method, RR: Ridge Regression CA-RR: our method; LSSVR: Suykens et al, IJCNN, 2001; KRR: An et al, CVPR, 2007; RFR: Liaw et al, R News, 2002; GPR: Chan et al, CVPR, 2008; RR: Chen et al, BMVC, 2012; Age Estimation Results Not based on regression What is OHRank? Proposed method, SVR: Support Vector Regression CA-SVR: our method; AGES: Geng et al, TPAMI, 2007; RUN: Yan et al, ICCV, 2007; Ranking: Yan et al, ICME, 2007; RED-SVM: Chang et al, ICPR, 2010; LARR: Guo et al, TIP, 2008; MTWGP: Zhang et al, CVPR, 2010; OHRank: Chang et al, CVPR, 2011; SVR: Guo et al, TIP, 2008; OHRank - Ordinal Hyperplanes Ranker Delta 0/1 function SVM score for older than k This is 104 slower than closed-form solution of regression Robustness Against Sparse and Unbalanced Data (Effects of removing random/certain label groups) Age Estimation Crowd Counting Feature Selection by Attributes Shape plays a more important role than texture for younger ages. Summary • Has a simple and neat idea • Exploits cumulative dependent nature of label space • Addresses sparse and unbalanced data problem Support Vector Regression Datasets