Efficient Image Search and Retrieval using Compact Binary Codes Rob Fergus (NYU) Antonio Torralba (MIT) Yair Weiss (Hebrew U.)

Efficient Image Search and Retrieval using Compact

Binary Codes

Rob Fergus (NYU)Antonio Torralba (MIT)Yair Weiss (Hebrew U.)

How can we search them, based on visual content?

Large scale image search

Internet contains many billions of images

The Challenge:– Need way of measuring similarity between images– Needs to scale to Internet

Existing approaches to Content-Based Image Retrieval

• Focus of scaling rather than understanding image• Variety of simple/hand-designed cues:– Color and/or Texture histograms, Shape, PCA, etc.

• Various distance metrics– Earth Movers Distance (Rubner et al. ‘98)

• Most recognition approaches slow (~1sec/image)

Our Approach

• Learn the metric from training data

DO BOTH TOGETHER

• Use compact binary codes for speed

Large scale image/video search• Representation must fit in memory (disk too slow)

• Facebook has ~10 billion images (1010)• PC has ~10 Gbytes of memory (1011 bits) Budget of 101 bits/image

• YouTube has ~ a trillion video frames (1012)• Big cluster of PCs has ~10 Tbytes (1014 bits) Budget of 102 bits/frame

Binary codes for images

• Want images with similar contentto have similar binary codes

• Use Hamming distance between codes– Number of bit flips– E.g.:

• Semantic Hashing [Salakhutdinov & Hinton, 2007]– Text documents

Ham_Dist(10001010,10001110)=1

Ham_Dist(10001010,11101110)=3

Semantic Hashing

Address Space

Semantically similar images

Query address

Semantic

HashFunction

Query Image

Binary code

Images in database

[Salakhutdinov & Hinton, 2007] for text documents

Quite differentto a (conventional)randomizing hash

Semantic Hashing

• Each image code is a memory address• Find neighbors by exploring Hamming

ball around query address Address Space

Query address

Images in database

ChooseCode length

Radius

• Lookup time is independentof # of data points

• Depends on radius of ball & length of code:

Code requirements

• Similar images Similar Codes• Very compact (<102 bits/image)• Fast to compute• Does NOT have to reconstruct image

Three approaches:1. Locality Sensitive Hashing (LSH)2. Boosting3. Restricted Boltzmann Machines (RBM’s)

Input Image representation: Gist vectors

• Pixels not a convenient representation• Use Gist descriptor instead (Oliva & Torralba,

2001)• 512 dimensions/image (real-valued 16,384 bits)• L2 distance btw. Gist vectors not bad substitute for

human perceptual distance

Oliva & Torralba, IJCV 2001

NO COLOR INFORMATION

1. Locality Sensitive Hashing• Gionis, A. & Indyk, P. & Motwani, R. (1999)• Take random projections of data• Quantize each projection with few bits

0

1

0

10

1

101

No learning involved

Gist descriptor

2. Boosting• Modified form of BoostSSC

[Shaknarovich, Viola & Darrell, 2003]• Positive examples are pairs of similar images• Negative examples are pairs of unrelated images

0

10

1

0 1

Learn threshold & dimension for each bit (weak classifier)

3. Restricted Boltzmann Machine (RBM)

Hidden units

Visible units

Symmetric weights

• Type of Deep Belief Network• Hinton & Salakhutdinov, Science 2006

SingleRBMlayer

• Attempts to reconstruct input at visible layer from activation of hidden layer

W

Multi-Layer RBM: non-linear dimensionality reduction

512

512w1

Input Gist vector (512 dimensions)

Layer 1

512

256w2

Layer 2

256

Nw3

Layer 3

Output binary code (N dimensions)

Linear units at first layer

Training RBM models

1st Phase: Pre-training

Unsupervised

Can use unlabeled data (unlimited quantity)

Learn parameters greedily per layer

Gets them to right ballpark

2nd Phase: Fine-tuning

Supervised

Requires labeled data(limited quantity)

Back propagate gradients of chosen error function

Moves parameters to local minimum

Greedy pre-training (Unsupervised)

512

512w1

Input Gist vector (512 real dimensions)

Layer 1

Greedy pre-training (Unsupervised)

Activations of hidden units from layer 1 (512 binary dimensions)

512

256w2

Layer 2

Greedy pre-training (Unsupervised)

Activations of hidden units from layer 2 (256 binary dimensions)

256

Nw3

Layer 3

Fine-tuning: back-propagation of Neighborhood Components Analysis objective

512

512

Input Gist vector (512 real dimensions)

Layer 1

512

256Layer 2

256

NLayer 3

Output binary code (N dimensions)

w1 + ∆ w1

w2 + ∆ w2

w3 + ∆w3 w3

w2

w1

Neighborhood Components Analysis• Goldberger, Roweis, Salakhutdinov & Hinton, NIPS 2004• Tries to preserve neighborhood structure of input space– Assumes this structure is given (will explain later)

Points in output space (coordinate is activation probability of unit)

Toy example with 2 classes & N=2 units at top of network:

Neighborhood Components Analysis• Adjust network parameters (weights and biases)

to move:– Points of SAME class closer

– Points of DIFFERENT class away

Neighborhood Components Analysis• Adjust network parameters (weights and biases)

to move:– Points of SAME class closer

– Points of DIFFERENT class away

Points close in input space (Gist) will be close in output code space

Simple Binarization Strategy

Set threshold- e.g. use median

0

1

0 1

Overall Query Scheme

Query Image

RBM

Compute Gist

Binary codeBinary code

Gist descriptor

Image 1

Semantic Hash

Retrieved images <1ms

~1ms (in Matlab)

<10μs

Retrieval Experiments

Test set 1: LabelMe

• 22,000 images (20,000 train | 2,000 test)• Ground truth segmentations for all• Can define ground truth distance btw. images

using these segmentations

Defining ground truth • Boosting and NCA back-propagation require

ground truth distance between images• Define this using labeled images from LabelMe

Defining ground truth • Pyramid Match (Lazebnik et al. 2006, Grauman & Darrell 2005)

Defining ground truth • Pyramid Match (Lazebnik et al. 2006, Grauman & Darrell 2005)

CarCar

Sky

Tree

Car

Road

Building

Car

Tree

Road

Building

CarCar

Sky

Tree

Car

Road

Building

Car

Tree

Road

Building

CarCar

Sky

Tree

Car

Road

Building

Car

Tree

Road

Building

Varying spatial resolution to capture approximate spatial correspondance

Examples of LabelMe retrieval• 12 closest neighbors under different distance metrics

LabelMe Retrieval

Size of retrieval set % o

f 50

true

nei

ghbo

rs in

retr

ieva

l set

0 2,000 10,000 20,0000

LabelMe Retrieval

Size of retrieval set % o

f 50

true

nei

ghbo

rs in

retr

ieva

l set

0 2,000 10,000 20,0000

Number of bits% o

f 50

true

nei

ghbo

rs in

firs

t 500

retr

ieve

d

Test set 2: Web images

• 12.9 million images• Collected from Internet• No labels, so use Euclidean distance between

Gist vectors as ground truth distance

Web images retrieval%

of 5

0 tr

ue n

eigh

bors

in re

trie

val s

et

Size of retrieval set

Web images retrieval

Size of retrieval set

% o

f 50

true

nei

ghbo

rs in

retr

ieva

l set

% o

f 50

true

nei

ghbo

rs in

retr

ieva

l set

Size of retrieval set

Examples of Web retrieval

• 12 neighbors using different distance metrics

Retrieval Timings

Summary

• Explored various approaches to learning binary codes for hashing-based retrieval– Very quick with performance comparable to complex

descriptors

• More recent work on binarization– Spectral Hashing (Weiss, Torralba, Fergus NIPS 2009)