Back

SVM-KNN: Discriminative Nearest Neighbor Classification for Visual Category Recognition

Haotong ZhangA. BergM. MaireJitendra Malik

article2006CVPR1,377 citations

Combines nearest-neighbor retrieval with local support vector machine training to enable scalable multiclass image classification using complex perceptual distance functions without the steep computational cost of full SVM training.

Listen

Visual category recognition faces severe bottlenecks when scaling to thousands of object classes and learning from sparse training data. Standard Nearest Neighbor (NN) classifiers naturally handle multiclass scenarios and flexible distance metrics, but suffer from high variance when training examples are scarce. Conversely, Support Vector Machines (SVMs) provide superior classification boundaries but become computationally intractable on large multiclass problems and require expensive pairwise distance computations. The article demonstrates and evaluates a hybrid approach, named SVM-KNN, which bridges these methods by identifying the closest neighbors of an unclassified sample and training a localized, multiclass SVM specifically on that small subset.

The framework mimics biological vision by applying a coarse-to-fine recognition strategy. It first shortlists candidates using an inexpensive crude distance, identifies the top K nearest neighbors using a domain-specific accurate distance metric, transforms those neighbor distances into a kernel matrix, and resolves fine-grained classification locally via a Directed Acyclic Graph SVM. This method was evaluated across four established benchmark datasets covering digit recognition (MNIST and USPS), surface texture classification (CUReT), and multi-object recognition (Caltech-101) using specialized shape and texture distance functions.

Across all evaluations, the hybrid method consistently matched or surpassed state-of-the-art accuracy while maintaining low computational overhead. On the USPS digit benchmark using complex tangent distances, the method achieved a 2.59% error rate—approaching human error rates of 2.5%—where full SVM training was computationally intractable. On the 61-category CUReT dataset, it lowered texture classification error to 1.73%, outperforming standard nearest-neighbor approaches (2.53%) while avoiding the severe computational cost of global pairwise models. On Caltech-101, it achieved 59.05% accuracy with 15 training images per class and 66.23% with 30 images, outperforming standalone NN baselines (40.98%) and global SVM approaches (56.40%).

These findings indicate that organizations can achieve state-of-the-art visual classification without incurring the prohibitive computational and infrastructure costs of training global multiclass models. The local learning design lowers computational complexity during training to zero and restricts expensive distance evaluations to a compact local neighborhood, offering a highly practical trade-off between speed and discrimination accuracy.

Engineering teams building high-category classification pipelines should implement local SVM classification over pruned nearest-neighbor candidates rather than attempting full global model retraining. Operational deployments should adopt two-stage distance shortlisting to maximize query throughput. Although the results provide high confidence across multiple visual domains, decision-makers should note that query latency depends on the complexity of the accurate distance function and the chosen neighbor count. Further work should focus on testing the framework on larger real-world datasets with thousands of unconstrained visual categories.

University of California Berkeleycomputational efficiencydirected acyclic graphsengineeringdiscriminative nearest neighbor classificationvisual category recognitionmulticlass SVM
Zhang et al (2006).pdf
  • Paper: Large Margin DAGs for Multiclass Classification, John Platt et al. (1999). Introduces the Directed Acyclic Graph Support Vector Machine (DAGSVM) architecture that SVM-KNN directly uses to efficiently resolve multiclass boundaries over localized candidate subsets.
  • Paper: Distance Metric Learning for Large Margin Nearest Neighbor Classification, Kilian Q. Weinberger et al. (2005). Establishes large-margin distance metric learning for nearest neighbor classification, providing foundational insight into combining margin objectives with neighborhood search.
  • Paper: Support-vector networks, Corinna Cortes et al. (1995). Presents the foundational support vector machine formulation and decomposition principles that underpin the discriminative local classifiers in SVM-KNN.
  • Paper: On the Algorithmic Implementation of Multiclass Kernel-based Vector Machines, Koby Crammer et al. (2002). Examines multiclass kernel machine optimization formulations and computational trade-offs that motivate localized, subset-based SVM alternatives.
  • Paper: Neighbourhood Components Analysis, Jacob Goldberger et al. (2004). Introduces Neighbourhood Components Analysis to learn distance metrics optimizing k-nearest neighbor classification, establishing core principles of neighborhood-based discriminative modeling.
  • Paper: Visual categorization with bags of keypoints, Gabriella Csurka et al. (2004). Demonstrates bag-of-keypoints visual representations paired with SVM classifiers, forming the baseline category recognition pipeline evaluated in SVM-KNN.
  • Paper: In Defense of One-Vs-All Classification, Ryan Rifkin et al. (2004). Provides a comprehensive analysis of multiclass reduction techniques versus all-pairs and graph-based strategies for Support Vector Machines.
  • Paper: Learning the Kernel Matrix with Semidefinite Programming, Gert R. G. Lanckriet et al. (2004). Establishes semidefinite programming methods for learning valid kernel matrices from pairwise data relationships.
Cover for SVM-KNN: Discriminative Nearest Neighbor Classification for Visual Category Recognition

Abstract

We consider visual category recognition in the framework of measuring similarities, or equivalently perceptual distances, to prototype examples of categories. This approach is quite flexible, and permits recognition based on color, texture, and particularly shape, in a homogeneous framework. While nearest neighbor classifiers are natural in this setting, they suffer from the problem of high variance (in bias-variance decomposition) in the case of limited sampling. Alternatively, one could use support vector machines but they involve time-consuming optimization and computation of pairwise distances.

We propose a hybrid of these two methods which deals naturally with the multiclass setting, has reasonable computational complexity both in training and at run time, and yields excellent results in practice. The basic idea is to find close neighbors to a query sample and train a local support vector machine that preserves the distance function on the collection of neighbors.

Our method can be applied to large, multiclass data sets for which it outperforms nearest neighbor and support vector machines, and remains efficient when the problem becomes intractable for support vector machines. A wide variety of distance functions can be used and our experiments show state-of-the-art performance on a number of benchmark data sets for shape and texture classification (MNIST, USPS, CUReT) and object recognition (Caltech-101). On Caltech-101 we achieved a correct classification rate of 59.05%(±0.56%) at 15 training images per class, and 66.23%(±0.48%) at 30 training images.

Topics

computational efficiencydirected acyclic graphsengineeringk-nn classifierlearning designsobject recognitionsupport vectorsvmtrademulti-class problems

Keywords

discriminative nearest neighbor classificationvisual category recognitionmulticlass SVMsupport vector machinesnearest neighbor searchlocal learninggeometric blur descriptorstexture histogramsshape matchingperceptual distances

Table of Contents

  • 1. Introduction
  • 2. SVM-KNN
  • 3. Shape and texture distances
  • 3.1. χ2 distance for texture
  • 3.2. Marginal distance for texture
  • 3.3. Tangent distance
  • 3.4. Shape context based distance
  • 3.5. Geometric blur based distance
  • 3.6. Kernelizing the distance
  • 4. Performance on benchmark data sets
  • 4.1. MNIST
  • 4.3. CUReT
  • 4.2. USPS
  • 4.4. Caltech-101
  • A. Algorithm A
  • B. Algorithm B
  • 5. Conclusion
  • References

Knowls

  1. Knowl 1 — SVM-KNN Classification Algorithm

    algorithm

    The SVM-KNN algorithm is a hybrid non-parametric classifier designed for multiclass visual recognition with complex distance functions. Given a query image, the algorithm first identifies a shortlist of candidates using a computationally cheap crude distance, refines the candidate set to the KK nearest neighbors using a more accurate (and potentially non-Euclidean or non-metric) perceptual distance, converts the local distance matrix to a positive semi-definite kernel, and trains a Directed Acyclic Graph Support Vector Machine (DAGSVM) locally on those KK neighbors to classify the query.

    Input: Query image xx, training dataset D={(xi,yi)}i=1nD = \{(x_i, y_i)\}_{i=1}^n, crude distance function dcruded_{\text{crude}}, accurate distance function daccud_{\text{accu}}, shortlist size KslK_{\text{sl}}, local neighborhood size KK
    Output: Predicted category label y∗y^*
    Compute crude distances dcrude(x,xi)d_{\text{crude}}(x, x_i) for all xi∈Dx_i \in D
    Select the KslK_{\text{sl}} nearest training examples to xx under dcruded_{\text{crude}} to form shortlist SS
    Compute accurate distances daccu(x,xj)d_{\text{accu}}(x, x_j) for all xj∈Sx_j \in S
    Select the KK nearest training examples from SS under daccud_{\text{accu}} to form local neighborhood NK(x)N_K(x)
    if all training samples in NK(x)N_K(x) have the same label cc then
        return cc
    end if
    Retrieve or compute pairwise accurate distance matrix DND_{N} for NK(x)∪{x}N_K(x) \cup \{x\}
    Construct kernel matrix KNK_{N} via KN(u,v)=12(DN(u,0)+DN(v,0)−DN(u,v))K_{N}(u, v) = \frac{1}{2}(D_{N}(u, 0) + D_{N}(v, 0) - D_{N}(u, v))
    Find smallest eigenvalue λmin⁡\lambda_{\min} of KNK_{N}
    if λmin⁡<0\lambda_{\min} < 0 then
        KN←KN+∣λmin⁡∣IK_{N} \leftarrow K_{N} + |\lambda_{\min}| I
    end if
    Train multiclass DAGSVM classifier on NK(x)N_K(x) using kernel KNK_{N}
    Classify query xx with the trained DAGSVM
    return predicted label y∗y^*

  2. Knowl 2 — Kernel Construction and Regularization from Arbitrary Perceptual Distances

    model/method

    Arbitrary perceptual distance functions d(x,y)d(x, y) (such as tangent distance, shape context distance, and geometric blur matching distance) are transformed into Mercer kernels for support vector machines through a three-step procedure:

    1. Distance-to-Kernel Transformation: A distance function d(x,y)d(x, y) is converted to a kernel matrix using the distance kernel trick: K(x,y)=12(d(x,0)+d(y,0)−d(x,y))K(x, y) = \frac{1}{2}(d(x, 0) + d(y, 0) - d(x, y)) where the choice of the reference point 00 does not affect SVM optimization.

    2. Symmetrization: If the distance measure is asymmetric (d(x,y)≠d(y,x)d(x, y) \neq d(y, x)), a symmetric distance is formed by: dsym(x,y)=d(x,y)+d(y,x)d_{\text{sym}}(x, y) = d(x, y) + d(y, x)

    3. Spectrum Shift for Positive Semi-Definiteness: When the underlying distance violates the triangle inequality, the resulting kernel matrix KK may have negative eigenvalues. To obtain a valid positive semi-definite kernel matrix K′K', the smallest eigenvalue λmin⁡(K)\lambda_{\min}(K) is computed. If λmin⁡(K)<0\lambda_{\min}(K) < 0, its absolute value is added to the diagonal: K′=K+∣λmin⁡(K)∣IK' = K + |\lambda_{\min}(K)| I This diagonal shift strengthens self-similarity without altering relative pairwise cross-similarities.

  3. Knowl 3 — Asymptotic Time Complexity of SVM-KNN versus Global DAGSVM

    theoretical result

    Let nn denote the number of training examples, KslK_{\text{sl}} the size of the shortlist, KK the number of nearest neighbors participating in the local support vector machine, CcrudeC_{\text{crude}} the computational cost of evaluating a crude distance (such as Euclidean distance), CaccuC_{\text{accu}} the computational cost of evaluating an accurate distance (such as tangent distance or geometric blur matching), and #SV\#\text{SV} the total number of support vectors in a global classifier.

    Classifier Training Time Complexity Query Time Complexity
    DAGSVM (Global) O(Caccun2)O(C_{\text{accu}} n^2) O(Caccu⋅#SV)O(C_{\text{accu}} \cdot \#\text{SV})
    SVM-KNN None O(Ccruden+Caccu(Ksl+K2))O(C_{\text{crude}} n + C_{\text{accu}}(K_{\text{sl}} + K^2))

    When nn is large or CaccuC_{\text{accu}} is high, global SVM training becomes intractable due to O(n2)O(n^2) pairwise accurate distance computations. SVM-KNN bypasses offline training and performs quick initial pruning via CcrudenC_{\text{crude}} n, keeping online local SVM training on KK samples tractable.

  4. Knowl 4 — Geometric Blur and Texture Distance Functions for Visual Object Recognition

    equation

    Two combined distance formulations between image ILI_L (with mm feature points) and image IRI_R (with nn feature points) are defined for object recognition:

    Algorithm A Distance (Shape matching with texture marginals): DA(IL→IR)=1m∑i=1mmin⁡j=1…n∥FiL−FjR∥22D^A(I_L \to I_R) = \frac{1}{m} \sum_{i=1}^m \min_{j=1\dots n} \|F_i^L - F_j^R\|_2^2 DA(IL,IR)=DA(IL→IR)+DA(IR→IL)+λ∑k=1nfilt∥hkL−hkR∥1D^A(I_L, I_R) = D^A(I_L \to I_R) + D^A(I_R \to I_L) + \lambda \sum_{k=1}^{n_{\text{filt}}} \|h_k^L - h_k^R\|_1 where FiLF_i^L and FjRF_j^R are normalized geometric blur descriptors (radius ∼70\sim 70 pixels) sampled at edge locations, hkLh_k^L and hkRh_k^R are L1L_1-normalized 1D histograms of filter response magnitudes for the kk-th filter in a Leung-Malik filter bank of size nfiltn_{\text{filt}}, and λ=18\lambda = \frac{1}{8}.

    Algorithm B Distance (Shape matching with 1st-order spatial distortion): DB(IL→IR)=1m∑i=1mmin⁡j=1…n(∥FiL−FjR∥22+λr0∥riL−rjR∥2)D^B(I_L \to I_R) = \frac{1}{m} \sum_{i=1}^m \min_{j=1\dots n} \left( \|F_i^L - F_j^R\|_2^2 + \frac{\lambda}{r_0} \|r_i^L - r_j^R\|_2 \right) DB(IL,IR)=DB(IL→IR)+DB(IR→IL)D^B(I_L, I_R) = D^B(I_L \to I_R) + D^B(I_R \to I_L) where riL,rjR∈R2r_i^L, r_j^R \in \mathbb{R}^2 denote pixel coordinates of feature points measured relative to image centers, r0=270r_0 = 270 is the average image dimension in pixels, geometric blur features are computed at a medium scale (radius ∼42\sim 42 pixels), and λ=14\lambda = \frac{1}{4}.

  5. Knowl 5 — Classification Accuracy on Caltech-101 Benchmark

    empirical result

    On the Caltech-101 dataset (101 object classes plus 1 background class, total 102 classes), classification performance was evaluated across different training set sizes per category using mean recognition rate per class averaged over 10 random train/test splits.

    Under Algorithm A (geometric blur + texture marginals, 15 training images per class):

    Method Correctness Rate (%)
    SVM-KNN (K=300K = 300) 59.08 ±\pm 0.37
    DAGSVM 56.40 ±\pm 0.36
    Nearest Neighbor (K=1K = 1) 40.98 ±\pm 0.47

    Under Algorithm B (geometric blur with first-order spatial distortion, DAGSVM classifier):

    Method 15 Training Examples/Class 30 Training Examples/Class
    Algorithm B 59.05 ±\pm 0.56% 66.23 ±\pm 0.48%
    Lazebnik, Schmid Ponce (2006) 56.40% 64.60 ±\pm 0.80%
    Berg (2005) 52.00% N/A
    Mutch Lowe (2006) 51.00% 56.00%
    Grauman Darrell (2006) 49.52% 58.23%
    Wang, Zhang Fei-Fei (2006) 44.00% 63.00%

    Under the evaluation protocol of Ommer and Buhmann, Algorithm B achieved a correctness rate of 63.0%63.0\%, compared to 57.8%57.8\% for their method.

  6. Knowl 6 — Digit Recognition Error Rates on USPS Dataset

    empirical result

    The USPS handwritten digit dataset consists of 7,291 training images and 2,007 test images of size 16×1616 \times 16 pixels (with a human error rate of 2.5%2.5\%). For tangent distance, images were pre-smoothed using a Gaussian filter with σ=0.75\sigma = 0.75.

    Classifier L2L_2 Distance Error Rate (%) Tangent Distance Error Rate (%)
    SVM-KNN 4.285 (K=10K = 10) 2.59 (K=8K = 8)
    Nearest Neighbor 5.530 (K=3K = 3) 2.89 (K=1K = 1)
    DAGSVM 4.400 Intractable
    HKNN 3.930 N/A

    With L2L_2 distance, SVM-KNN achieved comparable error rates to DAGSVM while training in a fraction of the time by restricting SVMs to K=10K=10 local samples per query. With tangent distance, global DAGSVM was computationally intractable due to O(n2)O(n^2) pairwise tangent distance evaluations, whereas SVM-KNN with K=8K=8 ran at near 1-NN speed and lowered the test error to 2.59%2.59\%.

  7. Knowl 7 — Digit Classification Error Rates on MNIST Dataset

    empirical result

    On the MNIST handwritten digit benchmark (60,000 training and 10,000 testing images of size 28×2828 \times 28), SVM-KNN was evaluated using standard L2L_2 distance across all 60,000 training images, and shape context distance on a limited subset of the first 10,000 training images (resized to 70×7070 \times 70, appearance terms omitted, evaluated using 10-fold cross-validation).

    Classifier L2L_2 Distance Error Rate (%) Shape Context (Limited) Error Rate (%)
    SVM-KNN 1.66 (K=80K = 80, Ksl≈800K_{\text{sl}} \approx 800) 1.67 ±\pm 0.49 (K=20K = 20)
    Nearest Neighbor 2.87 (K=3K = 3) 2.20 ±\pm 0.77 (K=1K = 1)

    For L2L_2 distance, setting the shortlist size Ksl≈10KK_{\text{sl}} \approx 10K captured the relevant neighbors with no degradation in accuracy compared to full search.

  8. Knowl 8 — Texture Classification Accuracy on CUReT Dataset

    empirical result

    The CUReT database evaluation used 61 real-world texture categories with 92 images per category photographed under varying illumination and viewing conditions. The dataset was split into 46 training and 46 test images per class and evaluated over 5 repetitions of 2-fold cross-validation using the Pearson χ2\chi^2 test statistic between texton histograms.

    Classifier Test Error Rate (%) Training Time / Complexity
    SVM-KNN (K=70K = 70) 1.73 ±\pm 0.24 Local SVM on 70 neighbors
    DAGSVM 1.75 ±\pm 0.25 1,830 binary SVMs (15,130 CPU sec)
    Nearest Neighbor (K=3K = 3) 2.53 ±\pm 0.28 Baseline texton NN

    SVM-KNN achieved lower error than both standard 3-NN (2.53%2.53\%) and global DAGSVM (1.75%1.75\%) while avoiding the computation of 1,830 global pairwise SVM classifiers.

Coverage note — None was omitted; all key algorithms, kernelization methods, mathematical distance equations, complexity comparisons, and experimental benchmark results (MNIST, USPS, CUReT, Caltech-101) are included.

References

  1. 1.S. Belongie, J. Malik, and J. Puzicha. Shape matching and object recognition using shape contexts. IEEE Trans. Pattern Anal. Mach. Intell., 24(4):509–522, 2002.
  2. 2.A. C. Berg. Shape Matching and Object Recognition. PhD thesis, Computer Science Division, University of California, Berkeley, 2005.
  3. 3.A. C. Berg, T. L. Berg, and J. Malik. Shape matching and object recognition using low distortion correspondence. In CVPR, 2005.
  4. 4.A. C. Berg and J. Malik. Geometric blur for template matching. In CVPR, pages 607–614, 2001.
  5. 5.P. Bickel and K. Doksum. Mathematical statistics, volume 1. Prentice Hall, 2nd edition, 2001.
  6. 6.L. Bottou and V. Vapnik. Local learning algorithms. Neural Computation, 4(6):888–900, 1992.
  7. 7.Bromley and Säckinger. Neural-network and K-nearestneighbor classifiers. Technical Report 11359-910819-16TM, AT&T, 1991.
  8. 8.T. M. Cover. Estimation by the nearest neighbor rule. IEEE Trans. on Information Theory, 14(1):50–55, 1968.
  9. 9.K. Crammer and Y. Singer. On the algorithmic implementation of multiclass kernel-based vector machines. J. Mach. Learn. Res., 2:265–292, 2002.
  10. 10.K. J. Dana, B. van Ginneken, S. K. Nayar, and J. J. Koenderink. Reflectance and texture of real-world surfaces. ACM Trans. Graph., 18(1):1–34, 1999.
  11. 11.C. Domeniconi and D. Gunopulos. Adaptive nearest neighbor classification using support vector machines. In NIPS, pages 665–672, 2001.
  12. 12.L. Fei-Fei, R. Fergus, and P. Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In IEEE CVPR Workshop of Generative Model Based Vision (WGMBV), 2004.
  13. 13.E. Goldmeier. Similarity in visually perceived forms. Psychological Issues, 8(1):1–134, 1972.
  14. 14.K. Grauman and T. Darrell. Discriminative classification with sets of image features. In ICCV, 2005.
  15. 15.K. Grauman and T. Darrell. Pyramid match kernels: Discriminative classification with sets of image features (version 2). Technical Report CSAIL-TR-2006-020, MIT, 2006.
  16. 16.T. Hastie and R. Tibshirani. Discriminant adaptive nearest neighbor classification. IEEE Trans. Pattern Anal. Mach. Intell., 18(6):607–616, 1996.
  17. 17.A. D. Holub, M. Welling, and P. Perona. Combining generative models and fisher kernels for object recognition. In ICCV, 2005.
  18. 18.S. Lazebnik, C. Schmid, and J. Ponce. Beyond bags of features: Spatial pyramid matching for recognizing natural scene categories. In CVPR, 2006.
  19. 19.Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, and L. D. Jackel. Backpropagation applied to handwritten zip code recognition. Neural Computation, 1(4):541–551, Winter 1989.
  20. 20.Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradientbased learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, November 1998.
  21. 21.Y. Lee, Y. Lin, and G. Wahba. Multicategory support vector machines, theory, and application to the classification of microarray data and satellite radiance data. Journal of the American Statistical Association, 99:67 – 81, 2004.
  22. 22.T. Leung and J. Malik. Representing and recognizing the visual appearance of materials using three-dimensional textons. Int. J. Comput. Vision, 43(1):29–44, 2001.
  23. 23.L. Levina. Statistical Issues in Texture Analysis. PhD thesis, Department of Statistics, University of California, Berkeley, 2002.
  24. 24.D. G. Lowe. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vision, 60(2):91–110, 2004.
  25. 25.G. Mori, S. Belongie, and J. Malik. Shape contexts enable efficient retrieval of similar shapes. In CVPR, volume 1, pages 723–730, 2001.
  26. 26.G. Mori and J. Malik. Estimating human body configurations using shape context matching. In European Conference on Computer Vision LNCS 2352, volume 3, pages 666–680, 2002.
  27. 27.J. Mutch and D. Lowe. Multiclass object recognition using sparse, localized features. In CVPR, 2006.
  28. 28.B. Ommer and J. M. Buhmann. Learning compositional categorization models. In ECCV, 2006.
  29. 29.E. Pekalska, P. Paclik, and R. P. W. Duin. A generalized kernel approach to dissimilarity-based classification. J. Mach. Learn. Res., 2:175–211, 2002.
  30. 30.J. C. Platt. Using analytic qp and sparseness to speed training of support vector machines. In NIPS, pages 557–563, Cambridge, MA, USA, 1999. MIT Press.
  31. 31.J. C. Platt, N. Cristianini, and J. Shawe-Taylor. Large margin DAGs for multiclass classification. In NIPS, pages 547–553, 1999.
  32. 32.E. Rosch. Natural categories. Cognitive Psychology, 4:328–350, 1973.
  33. 33.B. Schölkopf. The kernel trick for distances. In NIPS, pages 301–307, 2000.
  34. 34.T. Serre, L. Wolf, and T. Poggio. Object recognition with features inspired by visual cortex. In CVPR, 2005.
  35. 35.P. Simard, Y. LeCun, and J. S. Denker. Efficient pattern recognition using a new transformation distance. In NIPS, pages 50–58, San Francisco, CA, USA, 1993. Morgan Kaufmann Publishers Inc.
  36. 36.P. Simard, Y. LeCun, J. S. Denker, and B. Victorri. Transformation invariance in pattern recognition-tangent distance and tangent propagation. In Neural Networks: Tricks of the Trade, pages 239–274, London, UK, 1998. Springer-Verlag.
  37. 37.D. W. Thompson. On Growth and Form. Cambridge University Press, 1917.
  38. 38.S. Thorpe, D. Fize, and C. Marlot. Speed of processing in the human visual system. Nature, 381:520–522, June 1996.
  39. 39.A. B. Torralba, K. P. Murphy, and W. T. Freeman. Sharing features: Efficient boosting procedures for multiclass object detection. In CVPR, pages 762–769, 2004.
  40. 40.M. Varma and A. Zisserman. A statistical approach to texture classification from single images. International Journal of Computer Vision, 62(1–2):61–81, Apr. 2005.
  41. 41.P. Vincent and Y. Bengio. K-local hyperplane and convex distance nearest neighbor algorithms. In NIPS, pages 985–992, 2001.
  42. 42.G. Wang, Y. Zhang, and L. Fei-Fei. Using dependent regions for object categorization in a generative framework. In CVPR, 2006.

Citation

MLA
Hao Zhang, et al. “SVM-KNN: Discriminative Nearest Neighbor Classification for Visual Category Recognition”. 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition - Volume 2 (CVPR'06), vol. 2, 2006, pp. 2126–36, https://doi.org/10.1109/CVPR.2006.301.
APA
Hao Zhang, Berg, A. C., Maire, M., & Malik, J. (2006). SVM-KNN: Discriminative Nearest Neighbor Classification for Visual Category Recognition. 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition - Volume 2 (CVPR'06), 2, 2126–2136. https://doi.org/10.1109/CVPR.2006.301
Chicago
Hao Zhang, A. C. Berg, M. Maire, and J. Malik. 2006. “SVM-KNN: Discriminative Nearest Neighbor Classification for Visual Category Recognition”. 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition - Volume 2 (CVPR'06) 2: 2126–36. https://doi.org/10.1109/CVPR.2006.301.
Harvard
Hao Zhang et al. (2006) “SVM-KNN: Discriminative Nearest Neighbor Classification for Visual Category Recognition”, 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition - Volume 2 (CVPR'06). IEEE, pp. 2126–2136. Available at: https://doi.org/10.1109/CVPR.2006.301.
Vancouver
1. Hao Zhang, Berg AC, Maire M, Malik J (2006) SVM-KNN: Discriminative Nearest Neighbor Classification for Visual Category Recognition. In: 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition - Volume 2 (CVPR'06). IEEE, pp 2126–2136

BibTeX

@inproceedings{Hao_Zhang, title={SVM-KNN: Discriminative Nearest Neighbor Classification for Visual Category Recognition}, volume={2}, url={http://dx.doi.org/10.1109/CVPR.2006.301}, DOI={10.1109/cvpr.2006.301}, booktitle={2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition - Volume 2 (CVPR′06)}, publisher={IEEE}, author={Hao Zhang and Berg, A.C. and Maire, M. and Malik, J.}, pages={2126–2136} }
Metadata:Crossref

Access the Paper

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

Open PDF