Back

Iterative Quantization: A Procrustean Approach to Learning Binary Codes for Large-Scale Image Retrieval

Yunchao GongSvetlana LazebnikAlbert GordoFlorent Perronnin

article2013TPAMI1,720 citations

Proposes an alternating minimization method, Iterative Quantization, that finds an optimal orthogonal rotation to align dimensionally-reduced data with the vertices of a binary hypercube, significantly improving retrieval accuracy and scalability for large-scale image search.

Listen

Rapid growth in large-scale visual data requires efficient similarity search systems that minimize memory usage and processing latency without sacrificing search accuracy. Standard continuous feature representations are often too large to fit in operational memory at scale, while basic binary hashing techniques suffer from unbalanced data variance and severe quantization errors that degrade retrieval quality. The article addresses this operational bottleneck by introducing Iterative Quantization (ITQ), a method designed to learn compact, similarity-preserving binary codes by directly minimizing the quantization error when mapping projected data onto the vertices of a binary hypercube.

The authors evaluate ITQ across unsupervised, supervised, and nonlinear kernel settings using standard benchmark collections, including the 64,185-image CIFAR set, 580,000 Tiny Images, and a 1.2-million-image subset of ImageNet. The core approach applies dimensionality reduction—such as Principal Component Analysis (PCA) or supervised Canonical Correlation Analysis (CCA)—and then optimizes an orthogonal rotation matrix through an alternating minimization procedure. This rotation aligns the continuous feature embeddings with target binary hypercube vertices, balancing variance across code dimensions while maintaining linear scalability during training.

The analysis demonstrates several critical findings. First, ITQ consistently outperforms established hashing baselines; for instance, a 64-bit ITQ code matches the retrieval precision of a 256-bit unoptimized baseline, effectively delivering a fourfold reduction in code length. Second, integrating label information via CCA significantly boosts semantic precision, with ITQ-compressed supervised codes unexpectedly outperforming original, uncompressed continuous representations at sizes above 32 bits. Third, applying a nonlinear Random Fourier Feature mapping prior to quantization yields superior class label precision and allows codes to expand beyond the original feature dimensionality. Finally, testing on ImageNet shows that 950-bit binary visual attributes generated with ITQ achieve 38.98% retrieval precision (nearly identical to the uncompressed 32-bit floating-point performance of 39.24%) and improve novel category classification accuracy over continuous features when training data is scarce.

These findings indicate that organizations managing multi-million-image databases can dramatically lower server memory footprints and hardware costs by adopting ITQ-based binary compression without compromising search performance. ITQ-derived binary codes also enable direct training of fast linear classifiers, accelerating deployment cycles for novel visual categorization. When utilizing ITQ, decision-makers should pair the algorithm with supervised embeddings (such as CCA) if metadata is available, or use nonlinear kernel mappings when fine-grained or near-duplicate retrieval is required.

While ITQ provides substantial gains for exhaustive linear scans using Hamming distance, the article observes very low recall when binary codes are used for exact table-based hash lookups, indicating that hash-lookup indexing remains a challenge. The findings provide high confidence for linear-scan binary retrieval and attribute-based classification pipelines, but organizations should conduct pilot testing before deploying binary codes into exact hash indexing architectures.

Universitat Autònoma de BarcelonaUniversity of Illinois Urbana-ChampaignUniversity of North Carolina at Chapel HillXeroxbinary codescode lengthcorrelation analysisiterative quantizationlarge-scale image retrievalbinary codes
Gong et al (2013).pdf
Cover for Iterative Quantization: A Procrustean Approach to Learning Binary Codes for Large-Scale Image Retrieval

Abstract

This paper addresses the problem of learning similarity-preserving binary codes for efficient similarity search in large-scale image collections. We formulate this problem in terms of finding a rotation of zero-centered data so as to minimize the quantization error of mapping this data to the vertices of a zero-centered binary hypercube, and propose a simple and efficient alternating minimization algorithm to accomplish this task. This algorithm, dubbed iterative quantization (ITQ), has connections to multi-class spectral clustering and to the orthogonal Procrustes problem, and it can be used both with unsupervised data embeddings such as PCA and supervised embeddings such as canonical correlation analysis (CCA). The resulting binary codes significantly outperform several other state-of-the-art methods. We also show that further performance improvements can result from transforming the data with a nonlinear kernel mapping prior to PCA or CCA. Finally, we demonstrate an application of ITQ to learning binary attributes or “classemes” on the ImageNet dataset.

Topics

binary codescode lengthcorrelation analysisfourierhamming distancehashinghypercubeimage databaseprincipal component analysis (pca)quantizers

Keywords

iterative quantizationlarge-scale image retrievalbinary codesspectral hashingsemi-supervised hashingquantization errornearest neighbor searchproduct quantizationImageNet dataset

Table of Contents

  • 1 INTRODUCTION
  • 2 RELATED WORK
  • 3 UNSUPERVISED CODE LEARNING
  • 3.1 Unsupervised PCA Embedding
  • 3.2 Iterative Quantization
  • 4 EVALUATING UNSUPERVISED ITQ
  • 4.1 Datasets
  • 4.2 Protocols and Baseline Methods
  • 4.3 Results on CIFAR Dataset
  • 4.4 Results on 580,000 Tiny Images
  • 4.5 Evaluation of Hashing Performance
  • 5.2 Results
  • 6 ITQ WITH A KERNEL EMBEDDING
  • 6.1 Random Fourier Features
  • 6.2 Results
  • 7 LEARNING CLASSEMES FROM IMAGENET
  • 8 DISCUSSION
  • ACKNOWLEDGMENTS
  • REFERENCES

Knowls

  1. Knowl 1 — Iterative Quantization for Binary Code Learning

    algorithm

    Iterative Quantization (ITQ) is an alternating minimization algorithm designed to find an optimal orthogonal rotation matrix R∈Rc×cR \in \mathbb{R}^{c \times c} that minimizes the quantization error of mapping continuous projected data V∈Rn×cV \in \mathbb{R}^{n \times c} to binary hypercube vertices B∈{−1,1}n×cB \in \{-1, 1\}^{n \times c}.

    Input: Zero-centered projected data matrix V∈Rn×cV \in \mathbb{R}^{n \times c}, code length cc, number of iterations TT (default T=50T = 50).
    Output: Binary code matrix B∈{−1,1}n×cB \in \{-1, 1\}^{n \times c}, orthogonal rotation matrix R∈Rc×cR \in \mathbb{R}^{c \times c}.
    Initialize RR to a random c×cc \times c orthogonal matrix (obtained by QR decomposition of a random Gaussian matrix).
    for iter = 1 to TT do
        Update BB: B←sgn⁡(VR)B \leftarrow \operatorname{sgn}(V R)
        Compute SVD of the matrix BTVB^T V: BTV=SΩS^TB^T V = S \Omega \hat{S}^T
        Update RR: R←S^STR \leftarrow \hat{S} S^T
    end for
    B←sgn⁡(VR)B \leftarrow \operatorname{sgn}(V R)
    return B,RB, R

    Here, the element-wise sign function is defined as sgn⁡(v)=1\operatorname{sgn}(v) = 1 if v≥0v \ge 0 and −1-1 otherwise. At each iteration, the assignment step sgn⁡(VR)\operatorname{sgn}(VR) globally minimizes the loss for a fixed RR, and the Orthogonal Procrustes update via Singular Value Decomposition (SVD) globally minimizes the loss for a fixed BB. Consequently, the objective function value monotonically non-increases, guaranteeing convergence to a local minimum.

  2. Knowl 2 — Iterative Quantization Error Objective

    equation

    Given zero-centered data X∈Rn×dX \in \mathbb{R}^{n \times d} projected to a continuous cc-dimensional space V=XW∈Rn×cV = XW \in \mathbb{R}^{n \times c} using an embedding matrix W∈Rd×cW \in \mathbb{R}^{d \times c}, learning a similarity-preserving binary code corresponds to finding an orthogonal rotation matrix R∈Rc×cR \in \mathbb{R}^{c \times c} (RTR=IcR^T R = I_c) and a binary code matrix B∈{−1,1}n×cB \in \{-1, 1\}^{n \times c} that minimize the Frobenius norm quantization loss:

    Q(B,R)=∥B−VR∥F2=tr⁡((B−VR)(B−VR)T)\mathcal{Q}(B, R) = \|B - V R\|_F^2 = \operatorname{tr}\left((B - V R)(B - V R)^T\right)

    Expanding this objective yields:

    Q(B,R)=∥B∥F2+∥V∥F2−2tr⁡(BRTVT)=nc+∥V∥F2−2∑i=1n∑j=1cBijV~ij\mathcal{Q}(B, R) = \|B\|_F^2 + \|V\|_F^2 - 2\operatorname{tr}(B R^T V^T) = nc + \|V\|_F^2 - 2\sum_{i=1}^n \sum_{j=1}^c B_{ij} \tilde{V}_{ij}

    where V~=VR\tilde{V} = V R. Because the projected data VV and the constant ncnc are fixed, minimizing Q(B,R)\mathcal{Q}(B, R) is equivalent to maximizing tr⁡(BRTVT)\operatorname{tr}(B R^T V^T). Scaling the input feature vectors XX by a constant alters the value of Q(B,R)\mathcal{Q}(B, R) only by additive and multiplicative constants and leaves the optimal BB and RR unchanged.

  3. Knowl 3 — Unsupervised PCA Binary Code Embedding

    model/method

    In the unsupervised setting, ITQ generates the initial projection matrix W∈Rd×cW \in \mathbb{R}^{d \times c} by maximizing the variance of each binary bit and enforcing bit decorrelation under a continuous signed magnitude relaxation. For zero-centered data X∈Rn×dX \in \mathbb{R}^{n \times d} (satisfying ∑i=1nxi=0\sum_{i=1}^n x_i = 0), the variance maximization problem with orthogonal hyperplane constraints is formulated as:

    max⁡W∈Rd×c1ntr⁡(WTXTXW)subject toWTW=Ic\max_{W \in \mathbb{R}^{d \times c}} \frac{1}{n} \operatorname{tr}(W^T X^T X W) \quad \text{subject to} \quad W^T W = I_c

    The optimal matrix WW consists of the top cc eigenvectors of the data sample covariance matrix XTXX^T X. The resulting projected data V=XWV = XW is then rotated via the ITQ algorithm to obtain the final encoding transformation W~=WR\tilde{W} = WR, where each query x∈Rdx \in \mathbb{R}^d is encoded as h(x)=sgn⁡(xW~)h(x) = \operatorname{sgn}(x \tilde{W}).

  4. Knowl 4 — Supervised CCA Embedding for ITQ

    model/method

    When supervision is provided as continuous or binary label indicator vectors yi∈{0,1}ty_i \in \{0, 1\}^t forming a matrix Y∈{0,1}n×tY \in \{0, 1\}^{n \times t}, Canonical Correlation Analysis (CCA) is employed to learn a projection matrix that maximizes correlation between feature projections XwkX w_k and label projections YukY u_k.

    The projection vectors wk∈Rdw_k \in \mathbb{R}^d are obtained by solving the regularized generalized eigenvalue problem:

    XTY(YTY+ρIt)−1YTXwk=λk2(XTX+ρId)wkX^T Y (Y^T Y + \rho I_t)^{-1} Y^T X w_k = \lambda_k^2 (X^T X + \rho I_d) w_k

    where ρ>0\rho > 0 is a small regularization parameter (set to 10−410^{-4}).

    To construct the cc-dimensional embedding matrix W~∈Rd×c\tilde{W} \in \mathbb{R}^{d \times c}, the top cc generalized eigenvectors wkw_k are scaled by their corresponding eigenvalues λk\lambda_k (wk←λkwkw_k \leftarrow \lambda_k w_k) to emphasize the most discriminative directions and suppress noise from trailing eigenvectors. The embedded data V=XW~V = X\tilde{W} is subsequently quantized using ITQ.

  5. Knowl 5 — Nonlinear Kernel ITQ via Explicit Random Fourier Features

    model/method

    To capture nonlinear relationships and enable code lengths cc larger than the input feature dimension dd, data vectors x∈Rdx \in \mathbb{R}^d are mapped to a DD-dimensional explicit Random Fourier Feature (RFF) space to approximate a Gaussian kernel K(x,y)=exp⁡(−∥x−y∥22σ2)K(x, y) = \exp\left(-\frac{\|x - y\|^2}{2\sigma^2}\right).

    Each coordinate of the RFF mapping ΦD(x)=[Φw1,b1(x),…,ΦwD,bD(x)]T∈RD\Phi^D(x) = [\Phi_{w_1, b_1}(x), \dots, \Phi_{w_D, b_D}(x)]^T \in \mathbb{R}^D is computed as:

    Φwk,bk(x)=2cos⁡(xwk+bk)\Phi_{w_k, b_k}(x) = \sqrt{2} \cos(x w_k + b_k)

    where wk∼N(0,1σ2Id)w_k \sim \mathcal{N}\left(0, \frac{1}{\sigma^2} I_d\right) and bk∼Uniform⁡[0,2π]b_k \sim \operatorname{Uniform}[0, 2\pi]. Setting D=3000D = 3000, linear PCA or CCA is applied directly on the mapped data ΦD(X)∈Rn×D\Phi^D(X) \in \mathbb{R}^{n \times D} to reduce dimensionality to cc, followed by ITQ rotation and binarization.

  6. Knowl 6 — Binary Classemes Representation via ITQ

    model/method

    A binary descriptor for novel category retrieval and classification is constructed by training large banks of category classifiers and binarizing their continuous output scores using ITQ without intermediate dimensionality reduction.

    1. Base Classifiers: 950 base categories are sampled from the 1000 ILSVRC2010 classes. Linear SVM classifiers (C=2C=2) are trained on 4096-dimensional Fisher Vectors (composed of power- and L2L_2-normalized SIFT and color Fisher Vectors with 16 GMM components).
    2. Continuous Classeme Vector: For any image, the 950 real-valued decision values from the trained SVMs form a 950-dimensional feature vector.
    3. ITQ Binarization: The 950-dimensional continuous classemes are zero-centered across the dataset and rotated using the ITQ algorithm to obtain a compact 950-bit binary descriptor per image. Hamming distances or linear SVMs on top of these binary vectors are used for downstream zero-shot or novel category tasks.

  7. Knowl 7 — Empirical Retrieval Performance of Unsupervised PCA-ITQ

    empirical result

    On the CIFAR (64,185 images, 320-D GIST) and 580,000 Tiny Images datasets evaluated using Euclidean neighbor ground truth (average distance to 50th nearest neighbor):

    1. For small and medium code sizes (16 to 128 bits), PCA-ITQ achieves higher Mean Average Precision (mAP) than PCA with Random Rotation (PCA-RR), Non-orthogonal PCA relaxation (PCA-Nonorth), Spectral Hashing (SH), and Locality-Sensitive Hashing (LSH).
    2. Refinement via ITQ on a 64-bit code achieves Euclidean retrieval mAP comparable to a 256-bit PCA-RR baseline.
    3. When evaluating semantic consistency via average precision at top 500 retrieved images using CIFAR class labels, PCA-ITQ outperforms PCA-RR, PCA-Nonorth, SH, and Shift-Invariant Kernel Locality-Sensitive Hashing (SKLSH) across all bit rates from 16 to 256 bits.

  8. Knowl 8 — Retrieval Performance of Supervised CCA-ITQ

    empirical result

    Evaluating class label retrieval (top 500 average precision) on CIFAR:

    1. Supervised CCA-ITQ trained with clean class labels achieves the highest precision (e.g., ≈0.38\approx 0.38 at 16 bits to ≈0.44\approx 0.44 at 256 bits), substantially exceeding semi-supervised hashing (SSH-ITQ) and unsupervised PCA-ITQ.
    2. CCA-ITQ trained with noisy search keywords from 580,000 unverified Tiny Images maintains a large performance margin over unsupervised PCA-ITQ and SSH-ITQ.
    3. For code sizes greater than 32 bits, binarized CCA-ITQ codes outperform continuous, uncompressed CCA-projected data in top-500 semantic retrieval precision.

  9. Knowl 9 — Retrieval and Classification Performance of Binary Classemes on ILSVRC2010

    data/table

    On 50 held-out test categories of ILSVRC2010 (68,295 images total, 1,000 test queries), 950-bit binary classemes produced by ITQ preserve the retrieval precision of continuous 950-dimensional classemes (within ≈0.5%\approx 0.5\%) and outperform other binarization techniques. When training linear SVMs for novel category recognition with very small training sample sizes (1% and 5% of the training set), 950-bit Classeme-ITQ outperforms continuous classemes and 4096-dimensional continuous Fisher Vectors.

    Method Dimensionality Precision@50 (%)
    Fisher Vector 4096×324096 \times 32 bits 33.40
    Classeme 950×32950 \times 32 bits 39.24
    Classeme-Threshold 950×1950 \times 1 bits 31.54
    Classeme-ITQ 950×1950 \times 1 bits 38.98
    Classeme-RR 950×1950 \times 1 bits 37.06
    Classeme-SH 950×1950 \times 1 bits 22.66
    Classeme-LSH 950×1950 \times 1 bits 36.27
    Classeme-SKLSH 950×1950 \times 1 bits 33.35
    Method 1% 5% 10% 50% 100%
    Fisher Vector 35.55 52.21 57.11 66.21 69.16
    Classeme 38.54 51.49 56.18 64.31 66.77
    Classeme-Threshold 34.17 47.93 51.16 56.75 58.72
    Classeme-ITQ 40.58 53.32 56.62 61.12 62.68
    Classeme-RR 37.33 50.39 54.06 58.83 60.55

  10. Knowl 10 — Impact of Kernel Bandwidth on Locality Control in KPCA-ITQ

    empirical result

    In Random Fourier Feature KPCA-ITQ, the Gaussian kernel radius parameter σ\sigma directly controls the neighborhood radius of preserved similarity. When defining ground-truth neighbors at a small threshold δ\delta (equal to half the distance to the 50th nearest neighbor to capture near-duplicates):

    1. Kernel ITQ configured with σ=δ\sigma = \delta or σ=δ/2\sigma = \delta/2 achieves the highest precision-recall performance.
    2. Setting σ\sigma significantly larger (e.g., 2δ,4δ2\delta, 4\delta) or using linear PCA-ITQ fails to isolate near-duplicate points, resulting in poor retrieval performance.

Coverage note — None was omitted; all key theoretical formulations, alternating minimization algorithms, supervised/kernel extensions, application to visual classemes, and quantitative benchmark results were captured.

References

  1. 1.A. Andoni and P. Indyk. Near-optimal hashing algorithms for approximate nearest neighbor in high dimensions. Communication of the ACM, 2008.
  2. 2.A. Bergamo, L. Torresani, and A. Fitzgibbon. Picodes: Learning a compact code for novel-category recognition. NIPS, 2011.
  3. 3.M. B. Blaschko and C. H. Lampert. Correlational spectral clustering. CVPR, 2008.
  4. 4.P. Brasnett and M. Bober. Fast and robust image identification. ICPR, 2008.
  5. 5.M. Bronstein, A. Bronstein, N. Paragios, and F. Michel. Data fusion through cross-modality metric learning using similarity-sensitive hashing. CVPR, 2010.
  6. 6.O. Chapelle, J. Weston, and B. Schoelkopf. Cluster kernels for semi-supervised learning. NIPS, 2002.
  7. 7.O. Chum and J. Matas. Large scale discovery of spatilly related images. PAMI, 2010.
  8. 8.J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. CVPR, 2009.
  9. 9.R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and C.-J. Lin. Liblinear: A library for large linear classification. Journal of Machine Learning Research, 2008.
  10. 10.R. Fergus, A. Torralba, and Y. Weiss. Semi-supervised learning in gigantic image collections. NIPS, 2009.
  11. 11.D. P. Foster, R. Johnson, S. M. Kakade, and T. Zhang. Multi-view dimensionality reduction via canonical correlation analysis. Tech Report. Rutgers University, 2010.
  12. 12.J.-M. Frahm, P. Georgel, D. Gallup, T. Johnson, R. Raguram, C. Wu, Y.-H. Jen, E. Dunn, B. Clipp, S. Lazebnik, and M. Pollefeys. Building Rome on a cloudless day. In ECCV, 2010.
  13. 13.Y. Gong and S. Lazebnik. Iterative quantization: A Procrustean approach to learning binary codes. CVPR, 2011.
  14. 14.A. Gordo and F. Perronnin. Asymmetric distances for binary embeddings. CVPR, 2011.
  15. 15.J. He, R. Radhakrishnan, S.-F. Chang, and C. Bauer. Compact hashing with joint optimization of search accuracy and time. CVPR, 2011.
  16. 16.H. Hotelling. Relations between two sets of variables. Biometrika, 28:312–377, 1936.
  17. 17.H. Jégou, M. Douze, and C. Schmid. Hamming embedding and weak geometric consistency for large-scale image search. ECCV, 2008.
  18. 18.H. Jégou, M. Douze, and C. Schmid. Product quantization for nearest neighbor search. IEEE TPAMI, 2010.
  19. 19.H. Jégou, M. Douze, C. Schmid, and P. Pérez. Aggregating local descriptors into a compact image representation. CVPR, 2010.
  20. 20.A. Joly and O. Buisson. Random maximum margin hashing. CVPR, 2011.
  21. 21.A. Krizhevsky. Learning multiple layers of features from tiny images. Tech Report. University of Toronto, 2009.
  22. 22.B. Kulis and K. Grauman. Kernelized locality-sensitive hashing for scalable image search. In ICCV, 2009.
  23. 23.R.-S. Lin, D. Ross, and J. Yagnik. Spec hashing: Similarity preserving algorithm for entropy-based coding. CVPR, 2010.
  24. 24.W. Liu, S. Kumar, and S.-F. Chang. Hashing with graphs. ICML, 2011.
  25. 25.W. Liu, J. Wang, R. Ji, Y.-G. Jiang, and S.-F. Chang. Supervised hashing with kernels. CVPR, 2012.
  26. 26.D. G. Lowe. Distinctive image features from scale-invariant keypoints. IJCV, 2004.
  27. 27.S. Maji, A. C. Berg, and J. Malik. Classification using intersection kernel support vector machines is efficient. In CVPR, 2008.
  28. 28.D. Nister and H. Stewenius. Scalable recognition with a vocabulary tree. In CVPR, 2006.
  29. 29.M. Norouzi and D. J. Fleet. Minimal loss hashing for compact binary codes. ICML, 2011.
  30. 30.A. Oliva and A. Torralba. Modeling the shape of the scene: a holistic representation of the spatial envelope. IJCV, 2001.
  31. 31.L. Pauleve, H. Jegou, and L. Amsaleg. Locality sensitive hashing: a comparison of hash function types and querying mechanisms. Pattern Recognition Letters, 2010.
  32. 32.F. Perronnin and C. R. Dance. Fisher kernels on visual vocabularies for image categorization. CVPR, 2007.
  33. 33.F. Perronnin, Y. Liu, J. Sánchez, and H. Poirier. Large-scale image retrieval with compressed Fisher vectors. In CVPR, pages 3384–3391, 2010.
  34. 34.F. Perronnin, J. Sanchez, , and Y. Liu. Large-scale image categorization with explicit data embedding. CVPR, 2010.
  35. 35.F. Perronnin, J. Sanchez, and T. Mensink. Improving the fisher kernel for large-scale image classification. ECCV, 2010.
  36. 36.M. Raginsky and S. Lazebnik. Locality sensitive binary codes from sift-invariant kernels. NIPS, 2009.
  37. 37.A. Rahimi and B. Recht. Random features for large-scale kernel machines. NIPS, 2007.
  38. 38.N. Rasiwasia, P. Moreno, and N. Vasconcelos. Bridging the gap: Query by semantic example. IEEE Transactions on Multimedia, 2007.
  39. 39.R. Salakhutdinov and G. Hinton. Semantic hashing. International Journal of Approximate Reasoning, 2009.
  40. 40.B. Schölkopf, A. Smola, and K.-R. Müller. Kernel principal component analysis. In ICANN, 1997.
  41. 41.P. Schönemann. A generalized solution of the orthogonal procrustes problem. Psychometrika, 31, 1966.
  42. 42.G. Shakhnarovich, T. Darrell, and P. Indyk, editors. Nearest-Neighbors methods in Learning and Vision: Theory and Practice. MIT Press, 2006.
  43. 43.C. Strecha, A. M. Bronstein, M. M. Bronstein, and P. Fua. Ldahash: Improved matching with smaller descriptors. PAMI, 2010.
  44. 44.A. Torralba, R. Fergus, and W. Freenman. 80 million tiny images: a large dataset for non-parametric object and scene recognition. PAMI, 2008.
  45. 45.A. Torralba, R. Fergus, and Y. Weiss. Small codes and large image databases for recognition. CVPR, 2008.
  46. 46.L. Torresani, M. Szummer, , and A. Fitzgibbon. Efficient object category recognition using classemes. ECCV, 2010.
  47. 47.A. Vedaldi and A. Zisserman. Efficient additive kernels via explicit feature maps. CVPR, 2010.
  48. 48.G. Wang, D. Hoiem, and D. Forsyth. Learning image similarity from flickr groups using stochastic intersection kernel machines. ICCV, 2009.
  49. 49.J. Wang, S. Kumar, and S.-F. Chang. Semi-supervised hashing for scalable image retrieval. CVPR, 2010.
  50. 50.J. Wang, S. Kumar, and S.-F. Chang. Sequential projection learning for hashing with compact codes. ICML, 2010.
  51. 51.Y. Weiss, A. Torralba, and R. Fergus. Spectral hashing. NIPS, 2008.
  52. 52.S. X. Yu and J. Shi. Multiclass spectral clustering. ICCV, 2003.

Citation

MLA
Gong, Y., et al. “Iterative Quantization: A Procrustean Approach to Learning Binary Codes for Large-Scale Image Retrieval”. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 35, no. 12, 2013, pp. 2916–29, https://doi.org/10.1109/TPAMI.2012.193.
APA
Gong, Y., Lazebnik, S., Gordo, A., & Perronnin, F. (2013). Iterative Quantization: A Procrustean Approach to Learning Binary Codes for Large-Scale Image Retrieval. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(12), 2916–2929. https://doi.org/10.1109/TPAMI.2012.193
Chicago
Gong, Y., S. Lazebnik, A. Gordo, and F. Perronnin. 2013. “Iterative Quantization: A Procrustean Approach to Learning Binary Codes for Large-Scale Image Retrieval”. IEEE Transactions on Pattern Analysis and Machine Intelligence 35 (12): 2916–29. https://doi.org/10.1109/TPAMI.2012.193.
Harvard
Gong, Y. et al. (2013) “Iterative Quantization: A Procrustean Approach to Learning Binary Codes for Large-Scale Image Retrieval”, IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(12), pp. 2916–2929. Available at: https://doi.org/10.1109/TPAMI.2012.193.
Vancouver
1. Gong Y, Lazebnik S, Gordo A, Perronnin F (2013) Iterative Quantization: A Procrustean Approach to Learning Binary Codes for Large-Scale Image Retrieval. IEEE Transactions on Pattern Analysis and Machine Intelligence 35:2916–2929

BibTeX

@article{Gong_2013, title={Iterative Quantization: A Procrustean Approach to Learning Binary Codes for Large-Scale Image Retrieval}, volume={35}, ISSN={2160-9292}, url={http://dx.doi.org/10.1109/TPAMI.2012.193}, DOI={10.1109/tpami.2012.193}, number={12}, journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, publisher={Institute of Electrical and Electronics Engineers (IEEE)}, author={Gong, Yunchao and Lazebnik, Svetlana and Gordo, Albert and Perronnin, Florent}, year={2013}, month=Dec, pages={2916–2929} }
Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF