IEEE 2015 Conference on Computer Vision and Pattern Recognition Active Learning for Structured Probabilistic Models with Histogram Approximation Qing SunAnkit LaddhaDhruv Batra
2 Semantic segmentation Pose estimation Image classification No. Training images Accuracy
PASCAL VOC (~10,000 training images) 3
Leeds Sports Pose Dataset ( ~ 2000 annotated images) 4
5 ImageNet: Hired more than 25,000 AMT workers Human Hours ≈19 years Cost: $$$
Active Learning CRF Train Unlabeled Data Query Labeled Data Pick top n images to annotate 6 Peaky Low EntropyUniform High Entropy Compute “Informative”–ness Intractable!
Conditional Random Field 7 Node potential / Local Rewards Edge potential / Distributed Prior y1y1 y2y2 … ynyn yiyi kx kxk Gibbs distribution:
8 Goal Active Learning in Structured Output Models Contribution Novel Variational Approach for Entropy Estimation Challenge: Entropy Computation Intractable
Approximate Entropy via Sampling Gibbs sampling 9 Other modes not explored
Approximate Entropy via Variational Estimation 10 Simple distribution Family Q q*q* KL Divergence p All distributions
Approximate Entropy via Variational Estimation 11 Simple distribution Family Q q*q* Good Approximation p Inefficient Entropy Computation Inefficient Entropy Computation KL Divergence
Approximate Entropy via Variational Estimation 12 Simple distribution Family Q q*q* p Poor Approximation Efficient Entropy Computation Efficient Entropy Computation KL Divergence
Idea 2: Histogram Approximation 13 Idea 2: Histogram Approximation Approximation Histogram Idea 1: Delta Approximation Approximation Delta
14 Histogram Approximation Approximation Histogram
What is the optimal histogram? 15 Theorem [Sun, Laddha and Batra,CVPR 2015] Optimal Histogram Mass of Gibbs in Bins!
16 What is the optimal histogram?
Mass of Gibbs in a Bin P P P #P-complete
Maximum in Bins = Diverse Solutions 18 [3] Meier et al. The More the Merrier: Parameter Learning for Graphical Models with Multiple MAPs, ICML workshop, 2013 [2] Dhruv Batra et al. Diverse M-Best Solutions in Markov Random Fields, ECCV Top viewFront view
PDivMAP 19 Hamming distance terms are absorbed into the node terms Solve by reusing MAP solver! (graph cut, TRW-S, etc.) Lagrangian Relaxation: Primal: Hamming Distance CRF Score
Synthetic Experiment 20 Tree: Binary, 100 nodes, Potentials
21 True: Predicted:
Segmentation Experiments CRF 22
Baselines 23 SamplingVariational methodsLocal EntropyMargin-based Gibbs Sampling Perturb-and-MAP [Papandreou et al. ICCV 2011] Mean FieldMarginals [Luo et al. NIPS 2013] Min-Marginals [Kohli et al. CVIU 2008] [Batra et al. CVPR 2010] Margin-based [Roth et al. ECML 2006]
iCoseg Dataset [5] [4] Dhruv Batra, Adarsh Kowdle, Devi Parikh, Jiebo Luo, Tsuhan Chen, Interactively Co-segmentation Topically Related Images with Intelligent Scribble Guidance,. International Journal of Computer Vision (IJCV)
25 iCoseg Dataset Same performance with only 11% annotations +1%
CMU Geometric Context Dataset Support Vertical Sky Class Left Center Right Porous Solid 26 [6] Derek Hoiem,Alexei A. Efros, Martial Hebert, Recovering Surface Layout from an Image. IJCV, 75(1): (2007)
CMU Geometric Context Dataset 27 Same performance with only 14% annotations +3%
Summary Variational histogram approximiation for Active learning in structure models. Theoretically well motivated, easy to implement and outperforms baselines 28 Approximation Histogram 28
Thanks! 29 Qing SunAnkit LaddhaDhruv Batra machinelearning.ece.vt.edu Code is coming soon!