21
Efficient Inference for Fully-Connected CRFs with Stationarity Yimeng Zhang, Tsuhan Chen CVPR 2012

Efficient Inference for Fully-Connected CRFs with Stationarity

  • Upload
    kieve

  • View
    71

  • Download
    0

Embed Size (px)

DESCRIPTION

Efficient Inference for Fully-Connected CRFs with Stationarity. Yimeng Zhang, Tsuhan Chen CVPR 2012. Summary. Explore object-class segmentation with fully-connected CRF models Only restriction on pairwise terms is `spatial stationarity ’ (i.e. depend on relative locations) - PowerPoint PPT Presentation

Citation preview

Page 1: Efficient Inference for Fully-Connected CRFs with  Stationarity

Efficient Inference for Fully-Connected CRFs with Stationarity

Yimeng Zhang, Tsuhan ChenCVPR 2012

Page 2: Efficient Inference for Fully-Connected CRFs with  Stationarity

Summary

• Explore object-class segmentation with fully-connected CRF models

• Only restriction on pairwise terms is `spatial stationarity’ (i.e. depend on relative locations)

• Show how efficient inference can be achieved by– Using a QP formulation– Using FFT to calculate gradients in complexity

(linear in) O(NlogN)

Page 3: Efficient Inference for Fully-Connected CRFs with  Stationarity

Fully-connected CRF model

• General pairwise CRF model:

• Image I• Class labeling, X:• Label set, L: • V = set of pixels, N_i = neighbourhood of pixel i,

Z(I) = partition function, psi = potential functions

Page 4: Efficient Inference for Fully-Connected CRFs with  Stationarity

Fully-connected CRF model

• General pairwise CRF model:

• In fully-connected CRF, for all i, N_i = V

Page 5: Efficient Inference for Fully-Connected CRFs with  Stationarity

Unary Potential

• Unary potential generates a score for each object class per pixel (TextonBoost)

Page 6: Efficient Inference for Fully-Connected CRFs with  Stationarity

Pairwise Potential

• Pairwise potential measures compatibility of the labels at each pair of pixels

• Combines spatial and colour contrast factors

Page 7: Efficient Inference for Fully-Connected CRFs with  Stationarity

Pairwise Potential

• Colour contrast:

• Spatial term:

Page 8: Efficient Inference for Fully-Connected CRFs with  Stationarity

Pairwise Potential

• Learning the spatial term

Page 9: Efficient Inference for Fully-Connected CRFs with  Stationarity

MAP inference using QP relaxation• Introduce a binary indicator variable for each

pixel and label

• MAP inference expressed as a quadratic integer program, and relaxed to give the QP

Page 10: Efficient Inference for Fully-Connected CRFs with  Stationarity

MAP inference using QP relaxation

• QP relaxation has been proved to be tight in all cases (Ravikumar ICML 2006 [24])

• Moreover, it is convex whenever matrix of edge-weights is negative-definite

• Additive bound for non-convex case• QP requires O(KN) variables, LP requires (K^2E)

Page 11: Efficient Inference for Fully-Connected CRFs with  Stationarity

MAP inference using QP relaxation

• Gradient

• Derive fixed-point update by forming Lagrangian and setting its derivative to 0

Page 12: Efficient Inference for Fully-Connected CRFs with  Stationarity

Illustration of QP updates

Page 13: Efficient Inference for Fully-Connected CRFs with  Stationarity

Efficiently evaluating the gradient

• Required summation

• Would be a convolution without the color term• With color term is requires 5D-filtering• Can be approximated by clustering into C color

clusters, => C convolutions across

Page 14: Efficient Inference for Fully-Connected CRFs with  Stationarity

Efficiently evaluating the gradient

• Hence, for the case x_i = x_j, we need to evaluate

• Instead, evaluate for C clusters (C = 10 to 15)

• where

• Finally, interpolate

Page 15: Efficient Inference for Fully-Connected CRFs with  Stationarity

Update complexity• FFTs of each spatial filters can be calculated in

advance (K^2 filters)• At each update, we require C FFTs calculating,

O(CNlogN)• K^2 convolutions are needed, each requiring a

multiplication, O(K^2CN)• Terms can be added in Fourier domain, => only KC inverse FFTs needed, O(KCNlogN)

• Run-time per iteration < 0.1s for 213x320 pixels (+ downsampling by factor of 5)

Page 16: Efficient Inference for Fully-Connected CRFs with  Stationarity

MSRC synthetic experiment

• Unary terms randomized• Spatial distributions set to ground-truth

Page 17: Efficient Inference for Fully-Connected CRFs with  Stationarity

MSRC synthetic experiment

• Running times

Page 18: Efficient Inference for Fully-Connected CRFs with  Stationarity

Sowerby synthetic experiment

Page 19: Efficient Inference for Fully-Connected CRFs with  Stationarity

MSRC full experiment

• Use TextonBoost unary potentials• Compare with several other CRFs with same

unaries– Grid only– Grid + P^N (Kohli, CVPR 2008)– Grid + P^N + Cooccurrence (Ladickỳ, ECCV 2010)– Fully-connected + Gaussian spatial (Krähenbühl,

NIPS 2011)

Page 20: Efficient Inference for Fully-Connected CRFs with  Stationarity

MSRC full experiment

• Qualitative comparison

Page 21: Efficient Inference for Fully-Connected CRFs with  Stationarity

MSRC full experiment

• Quantitative comparison– Overall

– Per-class

– Timing: 2-8s per image