Diverse M-Best Solutions in Markov Random Fields , , , Dhruv Batra Payman Yadollahpour Abner Guzman-Rivera Greg Shakhnarovich TTI-Chicago / Virginia Tech TTI-Chicago UIUC TTI-Chicago.
Download ReportTranscript Diverse M-Best Solutions in Markov Random Fields , , , Dhruv Batra Payman Yadollahpour Abner Guzman-Rivera Greg Shakhnarovich TTI-Chicago / Virginia Tech TTI-Chicago UIUC TTI-Chicago.
Diverse M-Best Solutions in Markov Random Fields , , , Dhruv Batra Payman Yadollahpour Abner Guzman-Rivera Greg Shakhnarovich TTI-Chicago / Virginia Tech TTI-Chicago UIUC TTI-Chicago Local Ambiguity • Graphical Models Hat x1 x2 MAP … Inference xn Cat Most Likely Assignment (C) Dhruv Batra 2 Problems with MAP Model-Class is Wrong! -- Approximation Error Human Body ≠ Tree (C) Dhruv Batra Figure Courtesy: [Yang & Ramanan ICCV ‘11] 3 Problems with MAP Model-Class is Wrong! Not Enough Training Data! -- Approximation Error -- Estimation Error (C) Dhruv Batra 4 Problems with MAP Model-Class is Wrong! Not Enough Training Data! MAP is NP-Hard -- Approximation Error -- Estimation Error -- Optimization Error (C) Dhruv Batra 5 Problems with MAP Model-Class is Wrong! Not Enough Training Data! MAP is NP-Hard --Inherent ApproximationAmbiguity Error -- Estimation Error -- Optimization Error -- Bayes Error ? ? Rotating clockwise / Old Lady looking left / anti-clockwise? Young woman looking away? (C) Dhruv Batra One instance / Two instances? 6 Problems with MAP Model-Class is Wrong! Single Prediction = Uncertainty Mismanagement Not Enough Training Data! MAP is NP-Hard --Inherent Approximation Error Ambiguity -- Estimation Error -- Optimization Error Make Predictions! -- BayesMultiple Error (C) Dhruv Batra 7 Multiple Predictions xxx x xx xxx xxxx Sampling Porway & Zhu, 2011 TU & Zhu, 2002 Rich History (C) Dhruv Batra 8 Multiple Predictions Sampling M-Best MAP Porway & Zhu, 2011 TU & Zhu, 2002 Rich History Flerova et al., 2011 Fromer et al., 2009 Yanover et al., 2003 (C) Dhruv Batra Ideally: ✓ M-Best Modes 9 Multiple Predictions This Paper: Diverse M-Best in MRFs Sampling M-Best MAP - Porway & Zhu, 2011 TU & Zhu, 2002 Rich History (C) Dhruv Batra Don’t hope for diversity. Explicitly encode it. Ideally: Flerova et al., 2011 Fromer et Not guaranteed toal., be2009 modes. Yanover et al., 2003 ✓ M-Best Modes 10 MAP Integer Program kx1 (C) Dhruv Batra 11 MAP Integer Program 1 0 0 0 kx1 (C) Dhruv Batra 12 MAP Integer Program 0 1 0 0 kx1 (C) Dhruv Batra 13 MAP Integer Program 0 0 1 0 kx1 (C) Dhruv Batra 14 MAP Integer Program 0 0 0 1 kx1 (C) Dhruv Batra 15 MAP Integer Program 0 0 0 1 kx1 k2x1 (C) Dhruv Batra 16 MAP Integer Program 0 0 0 1 kx1 k2x1 (C) Dhruv Batra 17 MAP Integer Program Graphcuts, BP, Expansion, etc (C) Dhruv Batra 18 Diverse 2nd-Best Diversity MAP (C) Dhruv Batra 19 Diverse M-Best (C) Dhruv Batra 20 Diverse 2nd-Best Q1: How do we solve DivMBest? Q2: What kind of diversity functions are allowed? Q3: How much diversity? See Paper for Details (C) Dhruv Batra 21 Diverse 2nd-Best Diversity-Augmented Energy • Lagrangian Relaxation Many ways to solve: Primal 1. upergradient Ascent. Optimal. Slow. See Paper for Details 2. Binary Search. Optimal for M=2. Faster. 3. Grid-search on lambda. Sub-optimal. Fastest.Dualize Dual Div2Best energy Concave (Non-smooth) Lower-Bound on Div2Best En. (C) Dhruv Batra 22 Diverse 2nd-Best Q1: How do we solve Div2Best? Q2: What kind of diversity functions are allowed? Q3: How much diversity? See Paper for Details (C) Dhruv Batra 23 Diversity • [Special Case] 0-1 Diversity – M-Best MAP [Yanover NIPS03; Fromer NIPS09; Flerova Soft11] • [Special Case] Max Diversity [Park & Ramanan ICCV11] • Hamming Diversity • Cardinality Diversity See Paper for Details • Any Diversity (C) Dhruv Batra 24 Hamming Diversity 0 0 (C) Dhruv Batra 1 0 0 1 0 1 0 0 0 1 0 0 0 0 25 Hamming Diversity • Diversity Augmented Inference: (C) Dhruv Batra 26 Hamming Diversity • Diversity Augmented Inference: Unchanged. Can still use graph-cuts! Simply edit node-terms. Reuse MAP machinery! (C) Dhruv Batra 27 Experiments • 3 Applications – Interactive Segmentation: Hamming, Cardinality (in paper) – Pose Estimation: Hamming – Semantic Segmentation: Hamming • Baselines: – M-Best MAP – Confidence-Based Perturbation (No Diversity) (No Optimization) • Metrics – Oracle Accuracies • User-in-the-loop; Upper-Bound – Re-ranked Accuracies (C) Dhruv Batra 28 Experiment #1 • Interactive Segmentation – Model: Color/Texture + Potts Grid CRF – Inference: Graph-cuts – Dataset: 50 train/val/test images Image + Scribbles MAP 2nd Best MAP 1-2 Nodes Flipped (C) Dhruv Batra Diverse 2nd Best 100-500 Nodes Flipped 29 Experiment #1 96% +3.62% 95% 94% +1.61% 93% +0.05% 92% 91% 90% 89% MAP M-Best-MAP (Oracle) Confidence (Oracle) DivMBest (Oracle) M=6 (C) Dhruv Batra 30 Experiment #2 • Pose Tracking – Model: Mixture of Parts from [Park & Ramanan, ICCV ‘11] – Inference: Dynamic Programming – Dataset: 4 videos, 585 frames (C) Dhruv Batra Image Credit: [Yang & Ramanan, ICCV ‘11] 31 Experiment #2 • Pose Tracking w/ Chain CRF M Best Solutions (C) Dhruv Batra Image Credit: [Yang & Ramanan, ICCV ‘11] 32 Experiment #2 MAP (C) Dhruv Batra DivMBest + Viterbi 33 Experiment #2 85% Better DivMBest (Re-ranked) 80% PCP Accuracy 13% Gain 75% Same Features Same Model [Park & Ramanan, ICCV ‘11] (Re-ranked) 70% 65% 60% Confidence-based Perturbation (Re-ranked) 55% 50% 45% 1 51 101 151 201 251 301 #Solutions / Frame (C) Dhruv Batra 34 Experiment #3 • Semantic Segmentation – Model: Hierarchical CRF [Ladicky et al. ECCV ’10, ICCV ‘09] – Inference: Alpha-expansion – Dataset: Pascal Segmentation Challenge (VOC 2010) • 20 categories + background; 964 train/val/test images (C) Dhruv Batra Image Credit: [Ladicky et al. ECCV ’10, ICCV ’09] 35 Experiment #3 Input (C) Dhruv Batra MAP Best of 10-Div 36 Experiment #3 50% DivMBest (Oracle) Better PACAL Accuracy 45% 22%-gain possible 40% Same Features Same Model 35% DivMBest (Re-ranked) [Yadollahpour et al.] 30% Confidence-based Perturbation (Oracle) MAP 25% 20% 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 #Solutions / Image (C) Dhruv Batra 37 Summary • All models are wrong • Some beliefs are useful • DivMBest – First principled formulation for Diverse M-Best in MRFs – Efficient algorithm. Re-uses MAP machinery. – Big impact possible on many applications! (C) Dhruv Batra 38 Thank you! • Think about YOUR problem. • Are you or a loved one, tired of a single solution? • If yes, then DivMBest might be right for you!* * DivMBest is not suited for everyone. People with perfect models, and love of continuous variables should not use DivMBest. Consult your local optimization expert before starting DivMBest. Please do not drive or operate heavy machinery while on DivMBest. (C) Dhruv Batra 39