340 likes | 482 Views
Beyond MAP: Making Multiple Predictions: Diversity, DPPs and more. Dhruv Batra Virginia Tech. Alex Kulesza Univ. of Michigan. Dennis Park UC Irvine. Deva Ramanan UC Irvine. Schedule. Schedule. 1. Please interrupt & ask q uestions!. 2. All slides available online. Local Ambiguity.
E N D
Beyond MAP: Making Multiple Predictions: Diversity, DPPs and more. Dhruv BatraVirginia Tech Alex KuleszaUniv. of Michigan Dennis ParkUC Irvine Deva RamananUC Irvine
Schedule (C) Dhruv Batra
Schedule 1. Please interrupt & ask questions! 2. All slides available online. (C) Dhruv Batra
Local Ambiguity (C) Dhruv Batra slide credit: Fei-Fei Li, Rob Fergus & Antonio Torralba
Graphical Models to the Rescue! y1 MAP Inference y2 … Person xn Table Plate Most Likely Assignment (C) Dhruv Batra
Vision in 2000s (C) Dhruv Batra
Graphical Models in Vision Segmentation Object Recognition / Pose Estimation Left image Right image Disparity map Motion Flow Geometric Labelling Stereo Denoising (C) Dhruv Batra
Alpha-Expansion (C) Dhruv Batra Simulated Annealing
Problems Model-Class is Wrong! -- Approximation Error • Human Body ≠ Tree (C) Dhruv Batra Figure Courtesy: [Yang & Ramanan ICCV ‘11]
Problems Model-Class is Wrong! Not Enough Training Data! -- Approximation Error -- Estimation Error (C) Dhruv Batra
Problems Model-Class is Wrong! Not Enough Training Data! MAP is NP-Hard -- Approximation Error -- Estimation Error -- Optimization Error (C) Dhruv Batra
Biggest Problem Model-Class is Wrong! Not Enough Training Data! MAP is NP-Hard Inherent Ambiguity -- Approximation Error -- Estimation Error -- Optimization Error -- Bayes Error ? ? Rotating clockwise / anti-clockwise? Old Lady looking left / Young woman looking away? One instance / Two instances? (C) Dhruv Batra
Problems Single Prediction = Uncertainty Mismanagement Model-Class is Wrong! Not Enough Training Data! MAP is NP-Hard Inherent Ambiguity -- Approximation Error -- Estimation Error -- Optimization Error -- Bayes Error Make Multiple Predictions! (C) Dhruv Batra
Multiple Predictions (Diverse) M-Best MAP Dhruv10:45 – 11:30 Dennis9:30 – 10:15 Flerova et al., 2011 Fromeret al., 2009 Yanover et al., 2003 (C) Dhruv Batra
Multiple Predictions x x x x x x x x x x x x x Sampling Porway & Zhu, 2011 TU & Zhu, 2002 Rich History Deva1:30 – 2:00-pm (C) Dhruv Batra
Multiple Predictions DeterminentalPoint Process Alex2:00 – 3:153:45 – 4:30 Build a new model over sets that prefers diverse set (C) Dhruv Batra
Multiple Predictions DeterminentalPoint Process (Diverse) M-Best MAP Sampling Porway & Zhu, 2011 TU & Zhu, 2002 Rich History Alex2:00 – 3:153:45 – 4:30 Build a new model over sets that prefers diverse set Flerova et al., 2011 Fromeret al., 2009 Yanover et al., 2003 (C) Dhruv Batra
Multiple Predictions DeterminentalPoint Process (Diverse) M-Best MAP Sampling Porway & Zhu, 2011 TU & Zhu, 2002 Rich History Build a new model over sets that prefers diverse set Flerova et al., 2011 Fromeret al., 2009 Yanover et al., 2003 (C) Dhruv Batra
Schedule (C) Dhruv Batra All slides available online.
Notation andReview of CRFs (C) Dhruv Batra
Conditional Random Fields X1 • Discrete random variables • Factorized Model X2 1 1 10 0 kx1 10 0 … 10 10 Xi Xn 0 10 kxk Node Energies / Local Costs Edge Energies / Distributed Prior (C) Dhruv Batra
MAP Inference • In general NP-hard [Shimony ‘94] Approximate Inference • Heuristics: Loopy BP [Pearl, ‘88] • Greedy: α-Expansion [Boykov ’01, Komodakis ‘05] • LP Relaxations: [Schlesinger ‘76, Wainwright ’05, Sontag ’08, Batra ‘10] • QP/SDP Relaxations: [Ravikumar ’06, Kumar ‘09] (C) Dhruv Batra
MAP Integer Program kx1 (C) Dhruv Batra
MAP Integer Program 1 0 0 0 kx1 (C) Dhruv Batra
MAP Integer Program 0 1 0 0 kx1 (C) Dhruv Batra
MAP Integer Program 0 0 1 0 kx1 (C) Dhruv Batra
MAP Integer Program 0 0 0 1 kx1 (C) Dhruv Batra
MAP Integer Program 0 0 0 1 kx1 k2x1 (C) Dhruv Batra
MAP Integer Program 0 0 0 1 kx1 k2x1 (C) Dhruv Batra
MAP Integer Program Graphcuts, BP, Expansion, etc (C) Dhruv Batra
MAP Integer Program • LP view Graphcuts, BP, Expansion, etc MAP (C) Dhruv Batra
Schedule (C) Dhruv Batra All slides available online.