Small Codes and Large Image Databases for RecognitionCVPR 2008

Antonio Torralba, MITRob Fergus, NYU

Yair Weiss, Hebrew University

Outline

• Introduction• Methods• Experiment• Conclusion

Outline

• Introduction• Methods• Experiment• Conclusion

Summary

• Goal– efficient image search(real time on web-sized)

and fast, just require little memory, enable on standard hardware or handheld devices

• Approach– Use machine learning to convert Gist descriptor

to a compact binary code with a few hundred bits per image

Gist descriptor

• Global image representation• Describe the shapes occurring in an image

with one descriptor– Subdivide image in 4×4 sub images– Calculate Gabor responses in each of these– Create histograms of Gabor responses in each sub

image

Slide by James Hays and Alexei Efros

Gist descriptor

Slide by James Hays and Alexei Efros

Gist descriptor

• In this paper– 8 orientations ,4 frequency = 4×8×16 = 512

dimensional vector.– For smaller images (32×32 pixels), use 3 frequency

= 3×8×16 = 384 dimensions.

Binary Code

• Three reason– compression, it’s possible to represent images

with a very small number of bits and still maintain the information for recognition

Binary Code

– scaling up to web-size databases requires doing the calculations in memory. Fitting hundreds of millions of images into a few GB of memory means we have a budget of very few bytes per image.

– short binary codes allow very fast querying in standard hardware, either using hash tables or efficient bit-count operations

Locality Sensitive Hashing (LSH)

• high dimensional Euclidean space – finds nearest neighbors in constant time

• a number of random projections of that point into R1– each projection contributes a few bits

• when the number of bits is fixed and small– LSH can perform quite poorly

• In this paper– N = 30 bits

Outline

• Introduction• Methods• Experiment• Conclusion

Learning binary codes

• A database of images {xi}• a distance function D(i, j)• a binary feature vector yi = f(xi)• Hamming distance• N100(xi) - the 100 nearest neighbors of xi according to

the distance function D(i, j)• N100(yi) - the 100 descriptors yj that are closest to yi in

terms of Hamming distance• we would like N100(xi) = N100(yi) for all examples in

our training set

BoostSSC

• Boosting similarity sensitive coding• Learn original input space into a new space– distances between images can be computed using

a weighted Hamming distance.• Binary feature(M bits) – –

• weighted Hamming distance–

BoostSSC

• positive examples– pairs of images xi, xj , j ∈ N(xi).

• Negative examples– pairs of images that are not neighbors

• regression stump–

BoostSSC

• Minimize the square loss–

– K is the number of training pairs– Zk = 1, if the two images are neighbors; = −1,

otherwise–

• In this paper– – M around 30 bits

Restricted Boltzmann Machines

• Network of binary stochastic units•

– weights W, bias b

Hidden units: h

Symmetric weights: w

Visible units: v

Restricted Boltzmann Machines

• A probability can be assigned to a binary vector at the visible units –

• Convenient conditional distributions–

–

Learn weights and biases usingContrastive Divergence

Multi Layer RBM architecture‐

Training RBM models

• Pre training‐– Unsupervised– Use Contrastive Divergence to learn weights and biases– Gets parameters to right ballpark

• Fine tuning‐– Supervised– No longer stochastic– Backpropagate error to update parameters– Moves parameters to local minimum

Outline

• Introduction• Methods• Experiment• Conclusion

Two test datasets

• LabelMe– 22,000 images– Ground truth segmentations for all– Can define distance between images using these

segmentations• Web data[28]– 12.9 million images 32 × 32 colorimages – Subset of 80 million images– No labels, so use L2 distance between GIST vectors as

ground truth[28] A. Torralba, R. Fergus, and W. T. Freeman. Tiny images. Technical Report MIT-CSAIL-TR-2007-024, Computer Science and Artificial Intelligence Lab, Massachusetts Institute of Technology, 2007.

LabelMe retrieval

LabelMe retrieval

• what ground truth semantic similarity is– spatial pyramid matching over object labels

LabelMe retrieval

LabelMe retrieval

• On 2000 test images, N = 50

• • • • •

Web images retrieval

Web images retrieval

Retrieval speed evaluation

• Using multi-threading (M/T) on a quad-core

Pixel label

• On 2000 test images

Web images recognition

• On 2000 test images

Outline

• Introduction• Methods• Experiment• Conclusion

Conclusion

• Possible to build compact codes for retrieval– Fast and small on standard PC– Suitable for use on large database– Much room for improvement