Towards Semantic Embedding in Visual Vocabulary Towards Semantic Embedding in Visual Vocabulary The Twenty-Third IEEE Conference on Computer Vision and

Towards Semantic Embedding in Visual Vocabulary

Towards Semantic Embedding in Visual Vocabulary

The Twenty-Third IEEE Conference on Computer Vision and Pattern Recognition

Session: Object Recognition II: using Language, Tuesday 15, June 2010

Rongrong Ji1, Hongxun Yao1, Xiaoshuai Sun1, Bineng Zhong1, and Wen Gao1,2

1Visual Intelligence Laboratory, Harbin Institute of Technology2School of Electronic Engineering and Computer Science, Peking University

Overview

✴ Problem Statement

✴ Building patch-labeling correspondence

✴ Generative semantic embedding

✴ Experimental comparisons

✴ Conclusions

Overview





✴ Conclusions

Introduction

• Building visual codebooks– Quantization-based approaches

• K-Means, Vocabulary Tree, Approximate K-Means et al.

– Feature-indexing-based approaches• K-D Tree, R Tree, Ball Tree et al.

• Refining visual codebooks– Topic model decompositions

• pLSA, LDA, GMM et al.

– Spatial refinement• Visual pattern mining, Discriminative visual phrase et al.

Introduction

• With the prosperity of Web community

User Tags

Our Contribution

Introduction

• Problems to achieve this goal

– Supervision @ image level

– Correlative semantic labels

– Model generality

A traditional San Francisco street A traditional San Francisco street view with car and buildingsview with car and buildings

??Flower

Rose

Introduction

• Our contribution– Publish a ground truth path-labeling set

http://vilab.hit.edu.cn/~rrji/index_files/SemanticEmbedding.htm

– A generalized semantic embedding framework• Easy to deploy into different codebook models

– Modeling label correlations• Model correlative tagging in semantic embedding

Introduction

• The proposed framework

Build patch-level correspondenceSupervised visual codebook construction

Overview





✴ Conclusions

Building Patch-Labeling Correspondence

• Collecting over 60,000 Flickr photos– For each photo

• DoG detection + SIFT

• “Face” labels (for instance)


• Purify the path-labeling correspondences– Density-Diversity Estimation (DDE)– Formulation

• For a semantic label , its initial correspondence set is

– Patches extracted from images with label

– A correspondence from label to local patch

1 2

1 2, { , ,..., },

{ , ,..., }i

i i n ii

i i n i

d s d s d s

D s d d d s

le le le

1 2{ , ,..., }i nD d d d

ijleis

is jd

Purify le links in <Di, si>


• Density– For a given , Density reveals its

representability for :

• Diversity– For a given , Diversity reveals it

unique score for :

ld ldDen

is

21

1exp( || || )

l

m

d l j Lj

Den d dm

ldDivld

is

ln( )l

m md

i i

n nDiv

n n

Average neighborhood distance in m

neighborsnm: number of images in

neighborhood ni: number of total images with

label si

{ | }

. . j

j j j

Purifyi j d

d d d

D d DDE T

s t DDE Den Div

High T: Concerns more on Precision,

not Recall



• Case study: “Face” label (before DDE)


• Case study: “Face” label (after DDE)

Overview





✴ Conclusions

✴Generative Hidden Markov Random Field

✴ A Hidden Field for semantic modeling

✴ An Observed Field for local patch quantization

Generative Semantic Embedding

• Hidden Field– Each produces correspondence links (le) to

a subset of patch in the Observed Field– Links denotes semantic correlations

between and

1{ }mi iS s

is

ijl

is js

lij


WordNet::SimilarityPedersen et al. AAAI 04

• Observed Field– Any two nodes follow visual metric (e.g. L2)– Once there is a link between and , we

constrain by from the hidden field

1{ }ni iD d

ijle

jsjsid

id

sj

di

leij


• In the ideal case

– Each is conditionally independent given :

Neighbors in the Hidden Field

id

( | ) { ( | ) | ( | ) 0}i j i ji m

P D S P d s P d s

Feature set D is regarded as (partial) generative from the Hidden Field


– Formulize Clustering procedure

• Assign a unique cluster label to each

• Hence, D is quantized into a codebook with corresponding features

• Cost for codebook candidate C

P(C|D)=P(C)P(D|C)/P(D)

ic id

1{ }Kk kW w 1{ }Kk kV v


P(D) is a constraintSemantic Constraint

Visual Distortion

• Semantic Constraint P(C)– Define a MRF on the Observed Field

– For a quantizer assignments C, its probability can be expressed as a Gibbs distribution from the Hidden Field as


• That is– Two data points and in contribute to

if and only ifjdid kw

Observed FieldObserved Field

Hidden FieldHidden Field

ccii=c=cii

ssxx ssyy

ddjjddii

llxyxy


• Visual Constraint P(D|C)– Whether the codebook C is visually consistent

with current data D• Visual distortion in quantization


• Overall Cost• Finding MAP of P(C|D) can be converted into

maximizing its posterior probability

The visual constraint P(D|C)

The semantic constraints P(C)


• EM solution

• E step

• M step

Assign local patches to the closest clusters

Update the visual word center




• Model generation– To supervised codebook with label

independence assumption

– To unsupervised codebooks• Making all l=0

Overview





✴ Conclusions

Experimental Comparisons

Case study of ratios between inter-class distance and intra-class distance with and without semantic embedding in the Flickr dataset.


MAP@1 comparisons between our GSE model to Vocabulary Tree [1] and GNP [34] in Flickr 60,000 database.


Confusion table comparisons on PASCAL VOC dataset with method in [24].

Overview





✴ Conclusions

Conclusion

• Our contribution• Propagating Web Labeling from images to local

patches to supervise codebook quantization• Generalized semantic embedding framework for

supervised codebook building• Model correlative semantic labels in supervision

• Future works• Adapt one supervised codebook for different tasks• Move forward supervision into local feature

detection and extraction

Thank you!Thank you!

Documents

Towards Semantic Embedding in Visual Vocabulary Towards Semantic Embedding in Visual Vocabulary The Twenty-Third IEEE Conference on Computer Vision and