Multimedia Retrieval

Outline Overview Indexing Multimedia Generative Models & MMIR –Probabilistic Retrieval –Language models, GMMs Experiments –Corel experiments –TREC Video benchmark

Indexing Multimedia

A Wealth of Information Speech Audio Images Temporal composition Database

Associated Information id gender name country history picture Player Profile Biography

User Interaction Database Views results Gives examples video segments Poses a query query text feedback Evaluates

Indexing Multimedia Manually added descriptions –‘Metadata’ Analysis of associated data –Speech, captions, OCR, … Content-based retrieval –Approximate retrieval –Domain-specific techniques

Limitations of Metadata Vocabulary problem –Dark vs. somber Different people describe different aspects –Dark vs. evening

Limitations of Metadata Encoding Specificity Problem –A single person describes different aspects in different situations Many aspects of multimedia simply cannot be expressed unambiguously –Processes in left (analytic, verbal) vs. right brain (aesthetics, synthetic, nonverbal)

Approximate Retrieval Based on similarity –Find all objects that are similar to this one –Distance function –Representations capture some (syntactic) meaning of the object ‘Query by Example’ paradigm

N-dimensional space Feature extraction Ranking Display

Low-level Features

Query image

So, … Retrieval?!

IR is about satisfying vague information needs provided by users, (imprecisely specified in ambiguous natural language) by satisfying them approximately against information provided by authors (specified in the same ambiguous natural language) Smeaton

No ‘Exact’ Science! Evaluation is not done analytically, but experimentally –real users (specifying requests) –test collections (real document collections) –benchmarks (TREC: text retrieval conference) –Precision –Recall –...

Known Item

Query

Results …

Query

Results …

Semantic gap… raw multimedia data features concepts ?

Observation Automatic approaches are successful under two conditions: –the query example is derived from the same source as the target objects –a domain-specific detector is at hand

1. Generic Detectors

Retrieval Process Database Query Parsing Detector / Feature selection Filtering Ranking Query type Nouns Adjectives Camera operations People, Names Natural/physical objects Invariant color spaces

Parameterized detectors People detector Example Query text Topic 41 Results

Query Parsing Other examples of overhead zooming in views of canyons in the Western United States Nouns Adjectives The query type Find + names

Detectors Camera operations (pan, zoom, tilt, …) People (face based) Names (VideoOCR) Natural objects (color space selection) Physical objects (color space selection) Monologues (specifically designed) Press conferences (specifically designed) Interviews (specifically designed) Domain specific detectors FOCUSFOCUS The universe and everything

2. Domain knowledge

Player Segmentation Original image Initial segmentation Final segmentation

Show clips from tennis matches, starring Sampras, playing close to the net; Advanced Queries

3. Get to know your users

Mirror Approach Gather User’s Knowledge –Introduce semi-automatic processes for selection and combination of feature models Local Information –Relevance feedback from a user Global Information –Thesauri constructed from all users

N-dimensional space Feature extraction ClusteringRanking Concepts Display Thesauri

Low-level Features

Identify Groups

Representation Groups of feature vectors are conceptually equivalent to words in text retrieval So, techniques from text retrieval can now be applied to multimedia data as if these were text!

Query Formulation Clusters are internal representations, not suited for user interaction Use automatic query formulation based on global information (thesaurus) and local information (user feedback)

Interactive Query Process Select relevant clusters from thesaurus Search collection Improve results by adapting query –Remove clusters occuring in irrelevant images –Add clusters occuring in relevant images

Assign Semantics

Visual Thesaurus Correct cluster representing ‘Tree’, ‘Forest’ Glcm_47 ‘Incoherent’ cluster Fractal_23 Mis-labeled cluster Gabor_20

Learning Short-term: Adapt query to better reflect this user’s information need Long-term: Adapt thesaurus and clustering to improve system for all users

Thesaurus Only After Feedback

4. Nobody is unique!

Collaborative Filtering Also: social information filtering –Compare user judgments –Recommend differences between similar users People’s tastes are not randomly distributed You are what you buy (Amazon)

Collaborative Filtering Benefits over content-based approach –Overcomes problems with finding suitable features to represent e.g. art, music –Serendipity –Implicit mechanism for qualitative aspects like style Problems: large groups, broad domains

5. Ask for help

Query Articulation N-dimensional space Feature extraction How to articulate the query?

What is the query semantics ?

Details matter

Problem Statement Feature vectors capture ‘global’ aspects of the whole image Overall image characteristics dominate the feature-vectors Hypothesis: users are interested in details

Irrelevant Background QueryResult

Image Spots Image-spots articulate desired image details –Foreground/background colors –Colors forming ‘shapes’ –Enclosure of shapes by background colors Multi-spot queries define the spatial relations between a number of spots

Hist16HistSpotSpot+Hist Results Query Images

A: Simple Spot Query `Black sky’

B: Articulated Multi-Spot Query `Black sky’ above `Monochrome ground’

C: Histogram Search in `Black Sky’ images 2-4: 14:

Complicating Factors What are Good Feature Models? What are Good Ranking Functions? Queries are Subjective!

Probabilistic Approaches

Generative Models… A statistical model for generating data –Probability distribution over samples in a given ‘language’ M P ( | M )= P ( | M ) P ( | M, ) © Victor Lavrenko, Aug aka ‘Language Modelling’

Basic question: –What is the likelihood that this document is relevant to this query? P(rel|I,Q) = P(I,Q|rel)P(rel) / P(I,Q) … in Information Retrieval P(I,Q|rel) = P(Q|I,rel)P(I|rel)

‘Language Modelling’ Not just ‘English’ But also, the language of –author –newspaper –text document –image Hiemstra or Robertson? ‘Parsimonious language models explicitly address the relation between levels of language models that are typically used for smoothing.’

‘Language Modelling’ Guardian or Times?Not just ‘English’ But also, the language of –author –newspaper –text document –image

‘Language Modelling’ or ?Not just English! But also, the language of –author –newspaper –text document –image

Unigram and higher-order models Unigram Models N-gram Models Other Models –Grammar-based models, etc. –Mixture models = P ( ) P ( | ) P ( ) P ( ) P ( ) P ( ) P ( | ) P ( | ) P ( | ) © Victor Lavrenko, Aug. 2002

The fundamental problem Usually we don’t know the model M –But have a sample representative of that model First estimate a model from a sample Then compute the observation probability P ( | M ( ) ) M © Victor Lavrenko, Aug. 2002

Indexing: determine models Indexing –Estimate Gaussian Mixture Models from images using EM –Based on feature vector with colour, texture and position information from pixel blocks –Fixed number of components DocsModels

Retrieval: use query likelihood Query: Which of the models is most likely to generate these 24 samples?

Probabilistic Image Retrieval ?

Query Rank by P(Q|M) P(Q|M 1 ) P(Q|M 4 ) P(Q|M 3 ) P(Q|M 2 )

Topic Models Models P(Q|M 1 ) P(Q|M 4 ) P(Q|M 3 ) P(Q|M 2 ) Query P(D 1 |QM) P(D 2 |QM) P(D 3 |QM) P(D 4 |QM) Documents Query Model

Probabilistic Retrieval Model Text –Rank using probability of drawing query terms from document models Images –Rank using probability of drawing query blocks from document models Multi-modal –Rank using joint probability of drawing query samples from document models

Unigram Language Models (LM) –Urn metaphor Text Models P( ) ~ P ( ) P ( ) P ( ) P ( ) = 4/9 * 2/9 * 4/9 * 3/9 © Victor Lavrenko, Aug. 2002

Generative Models and IR Rank models (documents) by probability of generating the query Q: P( | ) = 4/9 * 2/9 * 4/9 * 3/9 = 96/9 P( | ) = 3/9 * 3/9 * 3/9 * 3/9 = 81/9 P( | ) = 2/9 * 3/9 * 2/9 * 4/9 = 48/9 P( | ) = 2/9 * 5/9 * 2/9 * 2/9 = 40/9

The Zero-frequency Problem Suppose some event not in our example –Model will assign zero probability to that event –And to any set of events involving the unseen event Happens frequently with language It is incorrect to infer zero probabilities –Especially when dealing with incomplete samples ?

Smoothing Idea: shift part of probability mass to unseen events Interpolation with background (General English) –Reflects expected frequency of events –Plays role of IDF – +(1- )

Hierarchical Language Model MNM Smoothed over multiple levels Alpha * P(T|Shot) + Beta * P(T|‘Scene’) + Gamma * P(T|Video) + (1–Alpha–Beta–Gamma) * P(T|Collection) Also common in XML retrieval –Element score smoothed with containing article

Image Models Urn metaphor not useful –Drawing pixels useless Pixels carry no semantics –Drawing pixel blocks not effective chances of drawing exact query blocks from document slim Use Gaussian Mixture Models (GMM) –Fixed number of Gaussian components/clusters/concepts

Key-frame representation Query model split colour channels Take samples Cr Cb Y DCT coefficients position EM algorithm

? Image Models Expectation-Maximisation (EM) algorithm –iteratively estimate component assignments re-estimate component parameters

Component 1Component 2Component 3 Expectation Maximization E M

animation Component 1Component 2Component 3 E M

VisualSEEk Querying by image regions and spatial laylout –Joint content-based/spatial querying –Automated region extraction –Direct indexing of color feature

Image Query Process

Color Similarity Color space similarity –measure the closeness in the HSV color space Color histograms distance –Minkowski metric between h q and h t Histogram quadratic distance –used in QBIC project –measure the weighted similarity between histograms –compute the cross similarity between colors It is computational expensive

Color sets –A compact alternative to color histograms Color set examples –Transform RGB to HSV Quantize the HSV color space to 2 hues, 2 saturations, and 2 values Assign a unique index m to each quantized HSV color => right dimensional binary space color set: a selection from the eight colors

Color sets and Back-Projection Process stage –Color selection 1, 2 colors, …, until extract salient regions –Back-projection onto the image map I[x,y] to the most similar color in the color set –Thresholding –Labeling

Color Set Query Strategy Given the query color set –perform several range queries on the query color set’s colors –take the intersection of these lists –minimize the sum of attributes in the intersection list

Single Region Query Region absolute location –fixed query location The Euclidean distance of centroids –bounded query location fall within a designed area: d q,t =0 otherwise: the Euclidean distance of the centroids

Index Structure Centroid location spatial access –Spatial quad-tree Rectangle (MBR) location spatial access –R-trees

Size Comparison Area –The absolute distance between two regions Spatial Extent –measure the distance among the width and the height of the MBRs –is much simpler than shape information

Single Region Query Strategy

Multiple Regions Query

Region Relative Location Use 2D-string

Spatial Invariance Provide scaling/rotation –approximate rotation invariance by providing different projections

Multiple Regions Query Strategy Relatives locations –perform query on all attributes except location –find the intersection of the region lists –for each candidate image the 2D-string is generated compared to the 2D-string of the query image

Query Formulation User tools –sketches regions –position them on the query grid –assign colors, size, and absolute location –may assign boundaries

Query Examples(1)

Query Examples(2)

Query Examples(3)

Bolbworld Image is treated as a few “blobs” –Image regions are roughly homogeneous with respect color and texture

Grouping Pixels into Regions For each pixel –assign a vector consisting of color, texture, and position features (8-D space) Model the distribution of pixels –Use EM algorithm to fit the mixture of Gaussians model to the data Perform spatial grouping –Connected pixels belonging to the same color/texture cluster

Describe the Regions For each region –store its color histogram Match the color of two regions –Use the quadratic distance between their histograms x and y –d 2 hist (x,y) = (x-y) T A(x-y) –Matrix A a symmetric matrix of weights between 0 and 1 represent the similarity between bin i and bin j

Querying in Blobworld The user composes a query –By submitting an image see its Blobworld representation select the relevant blobs to match specify the relative importance of the blob features –Atomic query Specify a particular blob with feature vector v i to match For each blob v j in the database image Find the distance between v i and v j –Measure the similarity between b i and b j –Take the maximum value

Querying in Blobworld Compound query –calculated using fuzzy-logic operator –user may specify a weight for each atomic query –rank the images according to the overall score indicate the blobs provided the highest score helps the user refine the query –Change the weighting of blob features –Specify new blobs to match

Indexing of Blobworld Indexing the color feature vectors –to speed up atomic queries –use R*-trees –higher dimensional data require larger data entries => lower fanout –Need a low dimensional approximation to the full color feature vectors use SVM to find A k : the best rank k approximation to the weight matrix A project the feature space into the subspace spanned by the rows of A k index the low-dimensional vector

Multimedia Cross-model Correlation Case study: automatic image captioning –Each image has a set of extracted regions –Some of the regions have a caption

Feature Extraction Discover image regions –Extracted by a standard segmentation algorithm –Each region is mapped into a 30-dim feature vector Mean, standard deviation of RGB values, average responses to various texture filters, position in the entire image layout, shape description How to capture cross-media correlations?

Mixed Media Graph (MMG) Graph construction –V(O): the vertex of object O –V(a i ): the vertex of the attribute value A= a i –Add an edge if and only if the two token values are close enough for each feature-vector, choose its k nearest neighbors => add the corresponding edges –the nearest neighbor relationship is not symmetric NN-links: the nearest neighbor links OAV-links: the object-attribute-value-links

Example of MMG

Correlation Discovery Correlation discovery by random walk –random walk with restart –To compute the affinity of node “B” for node “A” a random walker starts from node “A” choose randomly among the available edges every time –with probability c to go back to node “A” (restart) the steady-state probability that the random walker will find at node “B) –the affinity of “B” with respect to “A”

RWR Algorithm Given a query object O q –do an RWR from node q=V(O q ) –compute the steady state probability vector u q =(u q (1), …, u q (N)). Example –Given an image I 3 –Estimate u i3 for all nodes in the G MMG –Report the top few caption words for the image