PASCAL Slides

8/13/2019 PASCAL Slides

1/19

Discriminatively Trained Mixtures

of Deformable Part Models

Pedro Felzenszwalb and Ross Girshick

University of Chicago

David McAllesterToyota Technological Institute at Chicago

Deva RamananUC Irvine

http://www.cs.uchicago.edu/~pff/latent
http://www.cs.uchicago.edu/~pff/latenthttp://www.cs.uchicago.edu/~pff/latenthttp://www.cs.uchicago.edu/~pff/latenthttp://www.cs.uchicago.edu/~pff/latenthttp://www.cs.uchicago.edu/~pff/latenthttp://www.cs.uchicago.edu/~pff/latent


2/19

Model Overview

Mixture of deformable part models (pictorial structures) Each component has global template + deformable parts

Fully trained from bounding boxes alone


3/19

2 component bicycle model

root filterscoarse resolution

part filtersfiner resolution

deformationmodels


4/19

Object Hypothesis

Image pyramid HOG feature pyramid

Multiscale model captures features at two resolutions

Score of object hypothesisis sum of filter scores

minus deformation costs

Score of filter is dot

product of filter with HOG

features underneath it


5/19

Connection with linear classifier

concatenation of HOGfeatures and part

displacements and 0s

concatenation filters and

deformation parameters

root filter

part filter

def param

part filter

def param

...

root filter

part filter

def param

part filter

def param

...

w

}}w: model parametersz: latent variables: component label and

filter placements

score on detection windowxcan be written as

component 1

component 2


6/19

Latent SVM

Linear in w if z is fixed

Regularization Hinge loss


7/19

Latent SVM training

Non-convex optimization

Huge number of negative examples Convex if we fixzfor positiveexamples

Optimization:

- Initialize wand iterate:- Pick bestzfor each positive example- Optimize wvia gradient descent with data mining


8/19

Initializing w

For kcomponent mixture model:

Split examples into ksets based on bounding box aspect ratio

Learn kroot filters using standard SVM

- Training data: warped positive examples and randomwindows from negative images (Dalal & Triggs)

Initialize parts by selecting patches from root filters- Subwindows with strong coefficients- Interpolate to get higher resolution filters- Initialize spatial model using fixed spring constants


9/19

Car model

root filterscoarse resolution

part filtersfiner resolution

deformationmodels


10/19

Person model

root filters

coarse resolution

part filters

finer resolution

deformation

models


11/19

Bottle model

root filters

coarse resolution

part filters

finer resolution

deformation

models


12/19

Histogram of Gradient (HOG) features

Dalal & Triggs:

- Histogram gradient orientations in 8x8 pixel blocks (9 bins)- Normalize with respect to 4 different neighborhoods and truncate- 9 orientations * 4 normalizations = 36 features per block

PCA gives ~10 features that capture all information

- Fewer parameters, speeds up convolution, but costly projection at runtime Analytic projection: spans PCA subspace and easy to compute

- 9 orientations + 4 normalizations = 13 features

We also use 2*9 contrast sensitive features for 31 features total


13/19

Bounding box prediction

predict (x1, y1) and(x2, y2)from part locations

linear function trained using least-squares regression

(x1, y1) (x2, y2)


14/19

Context rescoring

Rescore a detection using context defined by all detections Let vibe the max score of detector for class iin the image

Let sbe the score of a particular detection

Let (x1,y1), (x2,y2)be normalized bounding box coordinates

f = (s, x1, y1, x2, y2, v1, v2... , v20)

Train class specific classifier

- f is positive example if true positive detection- f is negative example if false positive detection


15/19

Bicycle detection


16/19

More bicycles

False positives


17/19

Car


18/19

Person

Bottle Horse


19/19

Code

Source code for the system and modelstrained on PASCAL 2006, 2007 and 2008

data are available here:

http://www.cs.uchicago.edu/~pff/latent
http://www.cs.uchicago.edu/~pff/latenthttp://www.cs.uchicago.edu/~pff/latent

PASCAL Slides

Documents

Transcript of PASCAL Slides