learning spatial context: can stuff help us find things ?

Learning Spatial Context:

Can stuff help us find things?

Geremy HeitzDaphne Koller

April 14, 2008DAGS

Stuff (n): Material defined by a homogeneous or repetitive pattern of fine-scale properties, but has no specific or distinctive spatial extent or shape.

Thing (n): An object with a specific size and shape.

Outline

Sliding window object detection What is context? The Things and Stuff (TAS) model Results

Object Detection

Task: Find all the cars in this image. Return a “bounding

box” for each Evaluation:

Maximize true positives

Minimize false positives

Precision-Recall tradeoff

Sliding Window Detection Consider every bounding box

All shifts All scales Possibly all rotations

Each box gets a score: D(x,y,s,Θ)

Detections: Local peaks in D() Pros:

Covers the entire image Flexible to allow variety of

D()’s Cons:

Brute force – can be slow Only considers features in box

D = 1.5

D = -0.3

Features: Haar wavelets

Haar filters and integral imageViola and Jones, ICCV 2001

The average intensity in the block is computed with four sums independently of the block size.BOOSTING!

Features: Edge fragments

Weak detector = Match of edge chain(s) from training image to edgemap of test image

Opelt, Pinz, Zisserman, ECCV 2006

BOOSTING!

Histograms of oriented gradients

• Dalal & Trigs, 2006

• SIFT, D. Lowe, ICCV 1999

SVM!

Sliding Window Results

PASCALVisual Object Classes Challenge

Cows 2006

score(A,B) = |A∩B| / |AUB|True Pos: B s.t. score(A,B) > 0.5 for some AFalse Pos: B s.t. score(A,B) < 0.5 for all A False Neg: A s.t. score(A,B) < 0.5 for all B

B A

0.1 0.2 0.3 0.4 0.5

0.2

0.4

0.6

0.8

1

My DetectorINRIA-Douze

Recall Rate

Pre

cisi

on

Recall = TP / (TP + FN)Precision = TP / (TP + FP)

Satellite Detection Example

Quantitative Evaluation

0 40 80 120 160

0.2

0.4

0.6

0.8

1

False Positives Per Image

Rec

all R

ate

Why does this suck?

True Positives in Context

False Positives in Context

False Positives out of Context

Context!

What is Context?

What is Context?

Scene-Thing:

Stuff-Stuff:

gist car “likely”

keyboard “unlikely”

Thing-Thing:

Torralba et al., 2005

Gouldet al., 2008

Rabinovich et al., 2007

What is Context?

Stuff-Thing: Based on

intuitive “relationships”

Trees = no cars

Houses = cars nearby

Ro

ad =

cars here

Things

Candidate detections Bounding Box +

Score Boolean R.V. Ti

Ti = 1: Candidate is a positive detection

Thing-only model

Ti

ImageWindow

Wi

))(exp(1

1)(

WDWTP i

Stuff

Coherent image regions Coarse “superpixels” Feature vector Fj in Rn

Cluster label Sj

Stuff-only model Naïve Bayes

Sj

Fj

jjjjj SFPSPFSP )(),(

ssjj sSF ,~

Relationships

Descriptive Relations “Near”, “Above”,

“In front of”, etc. Choose a set R Rij: Relation

between detection i and region j

Relationship model

S72 = Trees

S 4 = H

ouses

S10 =

Ro

ad

Rij

TiSj

The TAS Model

RijTi Sj

Fj

ImageWindow

Wi

Wi: Window

Ti: Object Presence

Sj: Region Label

Fj: Region Features

Rij: Relationship

N

J

Unrolled Model

T1

S1

S2

S3

S4

S5

T2

T3

R21 = “Above”

R31 = “Left”

R13 = “In”

R33 = “In”

R11 = “Left”

CandidateWindows

ImageRegions

Learning Everything observed except Sj’s Expectation-Maximization

Mostly discrete variables Like Mixture-of-Gaussians

An ode to directed models:

Oh directed probabilistic modelsYou are so beautiful and palatableBecause unlike your undirected friendsYour parameters are so very interpretable

- Unknown Russian Mathematician

(Translated by Geremy Heitz)

RijTi Sj

Fj

ImageWindow

Wi

N

J

Learned Satellite Clusters

Inference

Goal:

Gibbs Sampling Easy to sample Ti’s given Sj’s

and vice versa

Could do distributional particles

RijTi Sj

Fj

ImageWindow

Wi

N

J

Results - Satellite

Prior:Detector Only

Posterior:TAS ModelRegion Labels

Results - Satellite

0 40 80 120 160

0.2

0.4

0.6

0.8

1

False Positives Per Image

Rec

all R

ate

Base DetectorTAS Model

PASCAL VOC Challenge

2005 Challenge 2232 images split into {train, val, test} Cars, Bikes, People, and Motorbikes

2006 5304 images plit into {train, test} 12 classes, we use Cows and Sheep

Results reported for challenge with state-of-the-art approaches Caveat: They didn’t get to see the test set

before the challenge, but I did!

Results – PASCAL

Cows

Results – PASCAL

Bicycles

Cluster #3

Results – PASCAL

Good examples

Discover “true positives”

Remove “false positives”

Results – VOC 2005

0.1 0.2 0.3 0.4 0.5

0.2

0.4

0.6

0.8

1

Recall Rate

Pre

cisi

on

TAS ModelBase DetectorsINRIA-Dalal

0.1 0.2 0.3 0.4 0.5 0.6

0.2

0.4

0.6

0.8

1

Recall Rate

Pre

cisi

on

Motorbike

Car

0.1 0.2 0.3 0.4

0.2

0.4

0.6

0.8

1

Recall Rate

Pre

cisi

on

Bicycle

0.1 0.2 0.3 0.4 0.5 0.6

0.2

0.4

0.6

0.8

1

Recall Rate

Pre

cisi

on

People

Results – VOC 2006

0.1 0.2 0.3 0.4 0.5

0.2

0.4

0.6

0.8

1TAS ModelBase DetectorsINRIA-Douze

Recall Rate

Pre

cisi

on

0.1 0.2 0.3 0.4 0.5

0.2

0.4

0.6

0.8

1

Recall Rate

Pre

cisi

on

Cow Sheep

Conclusions

Detectors can benefit from context The TAS model captures

an important type of context We can improve any sliding window

detector using TAS The TAS model can be interpreted and

matches our intuitions Geremy is smart

0 40 80 120 160

0.2

0.4

0.6

0.8

1

False Positives Per Image

Rec

all R

ate

Base DetectorTAS Model

Prior:Detector Only

Posterior:TAS Model

Region Labels

Detections in Context

Task: Identify all cars in the satellite image

Idea: The surrounding context adds info to the local window detector

+ =Houses

Road

Equations

))(exp(1

1)(

WDWTP i

jjjjj SFPSPFSP )(),(

ssjj sSF ,~