Back

What is the best multi-stage architecture for object recognition?

Kevin JarrettKoray KavukcuogluMarc'Aurelio RanzatoYann LeCun

article2009ICCV2,391 citations

Demonstrates that combining rectification, local contrast normalization, and pooling into multi-stage architectures is critical for high recognition accuracy on visual benchmarks, even enabling random, unlearned filters to achieve competitive performance.

Listen

Visual recognition systems often depend on multi-stage architectures that extract features through filter banks, non-linear processing, and pooling operations. However, system designers have faced uncertainty regarding the relative value of learned versus hardwired filters, the necessity of deep multi-stage hierarchies, and the specific impact of intermediate non-linear operations.

The article evaluates how non-linear transformations, filter training strategies (including random, unsupervised, and supervised schemes), and structural depth affect overall image recognition performance. It demonstrates the structural principles necessary to optimize feed-forward visual recognition architectures across benchmark tasks.

The researchers conducted systematic comparative experiments across multiple benchmark image datasets: Caltech-101 (object recognition with few training samples per class), NORB (3D object recognition with extensive pose and lighting variations across 24,300 training samples), and MNIST (handwritten digit classification with 60,000 training samples). The study benchmarked single-stage and two-stage architectures, testing combinations of non-linear rectifications (such as absolute value operations), local contrast normalization, spatial pooling methods, and different filter optimization schemes including Predictive Sparse Decomposition.

The analysis reveals five critical findings. First, incorporating non-linear rectification alongside local contrast normalization is the single most important factor for recognition accuracy; for example, on Caltech-101, adding rectification and normalization improved two-stage supervised accuracy from approximately 30% to 65.5%. Second, two-stage feature extraction architectures systematically outperform single-stage configurations, improving top accuracy on Caltech-101 from 54.2% to 65.5%. Third, surprisingly, a two-stage architecture using completely random, unlearned filters achieved a 62.9% recognition rate on Caltech-101 when paired with rectification and contrast normalization, showing that proper architecture naturally yields feature selectivity. Fourth, filter learning becomes essential when training data is plentiful; on the NORB dataset, supervised refinement significantly outperformed random filters as training sample size grew, ultimately reaching a low error rate of 5.6%. Fifth, combining unsupervised pre-training with supervised fine-tuning established a record low error rate of 0.53% on the unmodified MNIST benchmark.

These findings alter the prevailing understanding of neural feature extraction. Historical underperformance in supervised convolutional systems was not caused by over-parameterization on small datasets, but rather by inadequate non-linear processing. Proper architectural modules—especially rectification and local contrast normalization—prevent cancellation artifacts during spatial pooling and accelerate training convergence by standardizing feature variances. Consequently, architectural design choices can effectively offset data scarcity, while filter learning provides maximum advantage when larger labeled datasets are available.

Engineering teams developing computer vision pipelines should prioritize the integration of rectification and local contrast normalization across all feature extraction stages. For applications with limited labeled data, deploying properly regularized architectures with lightweight or feed-forward sparse initialization reduces the need for extensive manual feature engineering. When large labeled sets exist, end-to-end supervised fine-tuning remains the recommended approach to maximize accuracy.

The conclusions are drawn from controlled evaluations on single-family feature extractors across standard visual recognition benchmarks. While results demonstrate high consistency across grayscale object and digit datasets, real-world deployment across more complex visual domains, color imagery, or larger classification taxonomies may require additional validation.

New York Universityfeature extractionfilter banksobject recognitionFeatureLtersRecognition
Jarrett et al (2009).pdf
Cover for What is the best multi-stage architecture for object recognition?

Abstract

In many recent object recognition systems, feature extraction stages are generally composed of a filter bank, a non-linear transformation, and some sort of feature pooling layer. Most systems use only one stage of feature extraction in which the filters are hard-wired, or two stages where the filters in one or both stages are learned in supervised or unsupervised mode. This paper addresses three questions: 1. How does the non-linearities that follow the filter banks influence the recognition accuracy? 2. does learning the filter banks in an unsupervised or supervised manner improve the performance over random filters or hard-wired filters? 3. Is there any advantage to using an architecture with two stages of feature extraction, rather than one? We show that using non-linearities that include rectification and local contrast normalization is the single most important ingredient for good accuracy on object recognition benchmarks. We show that two stages of feature extraction yield better accuracy than one. Most surprisingly, we show that a two-stage system with random filters can yield almost 63% recognition rate on Caltech-101, provided that the proper non-linearities and pooling layers are used. Finally, we show that with supervised refinement, the system achieves state-of-the-art performance on NORB dataset (5.6%) and unsupervised pre-training followed by supervised refinement produces good accuracy on Caltech-101 (> 65%), and the lowest known error rate on the undistorted, unprocessed MNIST dataset (0.53%).

Topics

feature extractionfilter banksobject recognitionnon-linear transformations

Keywords

FeatureLtersRecognitionTrainingStageSupervisedObjectArchitectureObject RecognitionLter

Table of Contents

  • 1. Introduction
  • 1.1. Modules for dense feature extraction
  • 2. Model Architecture
  • 2.1. Combining Modules into a Hierarchy
  • 3. Training Protocol
  • 3.1. Unsupervised Training of Filter Banks using Predictive Sparse Decomposition
  • 4. Experiments
  • 4.1. NORB Dataset
  • 4.2. Random Filter Performance
  • 4.3. Handwritten Digits Recognition
  • 5. Conclusions
  • References

Knowls

  1. Knowl 1 — Multi-Stage Feature Extraction Module Architecture

    model/method

    A hierarchical feed-forward visual recognition architecture can be constructed by stacking one or two feature extraction stages, followed by a classifier (such as multinomial logistic regression or support vector machines). Each feature extraction stage is composed of up to four cascaded operations:

    1. Filter Bank Layer (FCSGF_{CSG}): Converts an input 3D array of size n1×n2×n3n_1 \times n_2 \times n_3 (n1n_1 feature maps of size n2×n3n_2 \times n_3) to an output 3D array yy of size m1×m2×m3m_1 \times m_2 \times m_3, where m1m_1 is the number of output feature maps and m2=n2−l1+1m_2 = n_2 - l_1 + 1, m3=n3−l2+1m_3 = n_3 - l_2 + 1 for 2D convolution filters of spatial kernel size l1×l2l_1 \times l_2. For input feature maps xix_i and filter kernels kijk_{ij} connecting input map xix_i to output map yjy_j, the layer computes: yj=gjtanh⁡(∑ikij∗xi)y_j = g_j \tanh\left(\sum_i k_{ij} * x_i\right) where ∗* denotes 2D discrete convolution, tanh⁡\tanh is the hyperbolic tangent non-linearity, and gjg_j is a trainable scalar gain coefficient.

    2. Rectification Layer (RabsR_{abs}): Applies a point-wise absolute value non-linearity across all components: yijk=∣xijk∣y_{ijk} = |x_{ijk}|. (Point-wise positive-part rectification max⁡(0,xijk)\max(0, x_{ijk}) produces comparable behavior).

    3. Local Contrast Normalization Layer (NN): Performs local subtractive and divisive normalization across spatial neighborhoods and across feature channels to implement local competition and variance equalization.

    4. Spatial Pooling and Subsampling Layer (PAP_A or PMP_M): Pools responses over a local spatial window of size p1×p2p_1 \times p_2 followed by downsampling by a factor SS. In Average Pooling (PAP_A), the output at each site is computed by a uniform boxcar averaging filter before downsampling. In Max-Pooling (PMP_M), the average is replaced by the maximum operator over non-overlapping spatial pooling windows.

  2. Knowl 2 — Local Contrast Normalization Formulation

    equation

    The Local Contrast Normalization module (NN) applies subtractive normalization followed by divisive normalization to an input tensor xx, where xijkx_{ijk} denotes the activation at feature map ii and spatial coordinates (j,k)(j, k):

    1. Subtractive Normalization: vijk=xijk−∑i,p,qwpqxi,j+p,k+qv_{ijk} = x_{ijk} - \sum_{i, p, q} w_{pq} x_{i, j+p, k+q} where wpqw_{pq} is a 2D Gaussian weighting window (e.g., 9×99 \times 9) normalized such that ∑i,p,qwpq=1\sum_{i, p, q} w_{pq} = 1.

    2. Divisive Normalization: yijk=vijkmax⁡(c,σjk)y_{ijk} = \frac{v_{ijk}}{\max(c, \sigma_{jk})} where the local standard deviation σjk\sigma_{jk} across feature channels and spatial neighbors is given by: σjk=(∑i,p,qwpqvi,j+p,k+q2)1/2\sigma_{jk} = \left( \sum_{i, p, q} w_{pq} v_{i, j+p, k+q}^2 \right)^{1/2} and cc is a constant per sample set to the mean over all spatial locations of σjk\sigma_{jk}, preventing division by near-zero values in uniform regions.

  3. Knowl 3 — Predictive Sparse Decomposition for Feed-Forward Filter Learning

    model/method

    Predictive Sparse Decomposition (PSD) is an unsupervised feature learning method that jointly learns a linear reconstruction dictionary and a fast feed-forward non-linear encoder. For an input patch XX, a reconstruction dictionary matrix WW (whose columns are dictionary elements), a sparse code ZZ, and a feed-forward parametric encoder C(X,K)=gtanh⁡(K∗X)C(X, K) = g \tanh(K * X) parameterized by filter coefficients KK, PSD optimizes the energy function:

    EPSD(X,Z,W,K)=∥X−WZ∥22+λ∥Z∥1+∥Z−C(X,K)∥22E_{PSD}(X, Z, W, K) = \|X - WZ\|_2^2 + \lambda \|Z\|_1 + \|Z - C(X, K)\|_{2}^2

    where λ\lambda is a sparsity-inducing regularization hyperparameter. The optimal sparse representation is defined as:

    Z∗=arg⁡min⁡ZEPSD(X,Z,W,K)Z^* = \arg\min_Z E_{PSD}(X, Z, W, K)

    Learning minimizes the empirical loss LPSD(W,K)=1P∑i=1Pmin⁡ZEPSD(Xi,Z,W,K)L_{PSD}(W, K) = \frac{1}{P} \sum_{i=1}^P \min_{Z} E_{PSD}(X^i, Z, W, K) over PP training patches via stochastic gradient descent, updating both the dictionary WW and the encoder parameters KK. Once learned, exact inference is bypassed during feed-forward recognition by directly computing the feature code as Z∗=C(X,K)Z^* = C(X, K), providing sparse, high-dimensional features deterministically without iterative optimization.

  4. Knowl 4 — Caltech-101 Benchmark Comparison of Architectures and Training Protocols

    data/table

    Recognition accuracy on the Caltech-101 benchmark evaluated across single-stage and two-stage architectures, comparing different filter initialization and training protocols: random features (R,RRR, RR), unsupervised PSD features (U,UUU, UU), random features with global supervised gradient refinement (R+,R+R+R^+, R^+R^+), unsupervised PSD features with global supervised gradient refinement (U+,U+U+U^+, U^+U^+), hand-crafted Gabor filters (GG), and Gabor filters with template initialization (GTGT). Classifiers are either multinomial logistic regression (log_reg) or Spatial Pyramid Match Kernel SVM (PMK-SVM). Results are average classification accuracy (%) over 101 categories plus background across 5 folds with 30 training samples per class:

    Single Stage System: [64.FCSG9×9−Pooling]−log_reg[64.F_{CSG}^{9\times 9} - \text{Pooling}] - \text{log\_reg}
    Protocol Rabs−N−PA5×5R_{abs}-N-P_A^{5\times 5} Rabs−PA5×5R_{abs}-P_A^{5\times 5} N−PM5×5N-P_M^{5\times 5} N−PA5×5N-P_A^{5\times 5} PA5×5P_A^{5\times 5}
    U+U^+ 54.2% 50.0% 44.3% 18.5% 14.5%
    R+R^+ 54.8% 47.0% 38.0% 16.3% 14.3%
    UU 52.2% 43.3% ( 1.6) 44.0% 17.2% 13.4%
    RR 53.3% 31.7% 32.1% 15.3% 12.1% ( 2.2)
    GG 52.3% - - - -
    Two Stage System: [64.FCSG9×9−Pooling]−[256.FCSG9×9−Pooling]−log_reg[64.F_{CSG}^{9\times 9} - \text{Pooling}] - [256.F_{CSG}^{9\times 9} - \text{Pooling}] - \text{log\_reg}
    Protocol Rabs−N−PA4×4R_{abs}-N-P_A^{4\times 4} Rabs−PA4×4R_{abs}-P_A^{4\times 4} N−PM4×4N-P_M^{4\times 4} N−PA4×4N-P_A^{4\times 4} PA4×4P_A^{4\times 4}
    U+U+U^+U^+ 65.5% 60.5% 61.0% 34.0% 32.0%
    R+R+R^+R^+ 64.7% 59.5% 60.0% 31.0% 29.7%
    UUUU 63.7% 46.7% 56.0% 23.1% 9.1%
    RRRR 62.9% 33.7% ( 1.5) 37.6% ( 1.9) 19.6% 8.8%
    GTGT 55.8% - - - -
    PMK-SVM Classifiers
    Single Stage: [64.FCSG9×9−Rabs/N/PA5×5]−PMK-SVM:U=64.0%[64.F_{CSG}^{9\times 9} - R_{abs}/N/P_A^{5\times 5}] - \text{PMK-SVM}: U = 64.0\%
    Two Stages: [64.FCSG9×9−Rabs/N/PA5×5]−[256.FCSG9×9−Rabs/N]−PMK-SVM:UU=52.8%[64.F_{CSG}^{9\times 9} - R_{abs}/N/P_A^{5\times 5}] - [256.F_{CSG}^{9\times 9} - R_{abs}/N] - \text{PMK-SVM}: UU = 52.8\%

    The table demonstrates three primary findings: (1) two feature extraction stages systematically outperform single-stage counterparts by ≈10–11%\approx 10\text{--}11\%; (2) random filters combined with rectification and normalization (Rabs−N−PAR_{abs}-N-P_A) achieve 62.9%62.9\% (two-stage), approaching learned filters (63.7%63.7\% for UUUU and 65.5%65.5\% for U+U+U^+U^+); (3) omitting rectification in average-pooling systems causes catastrophic performance degradation (8.8–32.0%8.8\text{--}32.0\%).

  5. Knowl 5 — Role of Rectification and Normalization Against Average-Pooling Signal Cancellation

    theoretical result

    In convolutional feature extraction hierarchies, absolute-value rectification (RabsR_{abs}) or positive-part rectification is necessary before average pooling (PAP_A) because linear or squashed filter outputs of alternating polarities within a spatial neighborhood cancel each other out during boxcar summation, propagating background noise rather than feature energy.

    Max-pooling (PMP_M) partially mitigates this requirement without explicit absolute-value rectification because the max operator selects dominant positive activations, preventing destructive interference. Local contrast normalization (NN) further enhances classification performance and accelerates gradient-based supervised optimization by equalizing activation variances across spatial locations and feature channels, behaving similarly to whitening/decorrelation operations.

  6. Knowl 6 — Data-Regime Dependence of Random vs. Learned Features on NORB

    empirical result

    On the 5-class Small NORB dataset (24,300 training and 24,300 test images), the performance gap between random filters (RRRR) and globally supervised learned filters (R+R+R^+R^+) strongly depends on the number of labeled training samples per class:

    • In small sample regimes (20 to 100 samples per class, matching the sample scarcity of Caltech-101), fixed random filter networks with rectification and normalization (FCSG−Rabs−N−PAF_{CSG}-R_{abs}-N-P_A) achieve test error rates comparable to supervised networks.
    • As labeled samples increase toward the full dataset size (4,860 samples per class), the test error of random filters plateaus and remains high, whereas supervised learning (R+R+R^+R^+) improves error monotonically.
    • On the full dataset, the FCSG−Rabs−N−PAF_{CSG}-R_{abs}-N-P_A architecture with R+R+R^+R^+ supervised learning achieves an error rate of 5.6%5.6\%, compared to 6.9%6.9\% for the same network trained without rectification and normalization (FCSG−PAF_{CSG}-P_A), confirming that architectural design and training scale complement each other.

  7. Knowl 7 — Spontaneous Orientation Selectivity in Random Filter Stages

    empirical result

    When synthesizing the optimal input pattern that maximizes the activation of a complex cell (a unit after the PAP_A pooling layer) in an FCSG−Rabs−N−PAF_{CSG}-R_{abs}-N-P_A stage initialized with purely random filter weights via gradient ascent on the pixel inputs, the resulting input stimuli are oriented, periodic spatial gratings.

    Slight statistical asymmetries present in random 2D filter kernels, when cascaded with point-wise rectification (RabsR_{abs}), local contrast normalization (NN), and spatial pooling (PAP_A), spontaneously create orientation-selective and phase-invariant complex receptive fields analogous to those produced by trained or hand-crafted Gabor filter stages.

  8. Knowl 8 — MNIST Digit Recognition via Two-Stage PSD Pre-training and Supervised Refinement

    empirical result

    A two-stage convolutional hierarchy evaluated on the standard unprocessed MNIST dataset (60,000 training and 10,000 test images of 28×2828 \times 28 grayscale digits) achieves a test error rate of 0.53%0.53\% without data augmentation or image distortions.

    The network architecture consists of:

    1. Stage 1: 32 feature maps with 5×55 \times 5 filters (FCSGF_{CSG}), absolute rectification (RabsR_{abs}), local contrast normalization (NN), and 2×22 \times 2 average pooling (PAP_A).
    2. Stage 2: 64 feature maps, each connected to 16 randomly selected stage-1 feature maps using 5×55 \times 5 filters (1,024 filters total), followed by RabsR_{abs}, NN, and 2×22 \times 2 average pooling (PAP_A).
    3. Classifier: Fully connected 2-layer neural network with 200 hidden units and 10 softmax output units trained with cross-entropy loss.

    Feature extractors are pre-trained stage-wise in unsupervised mode using Predictive Sparse Decomposition (with sparsity hyperparameter λ=0.2\lambda = 0.2 tuned on a 10,000-sample validation set). The full system is then refined end-to-end via stochastic gradient descent (the U+U+R+R+U^+U^+R^+R^+ protocol) for 30 epochs on 50,000 training samples and 3 additional epochs on the entire 60,000 training set.

Coverage note — None was omitted; all key architectural components, equations, empirical findings across Caltech-101, NORB, and MNIST, theoretical mechanisms for rectification/pooling, and random filter analyses are fully covered.

References

  1. 1.http://yann.lecun.com/exdb/mnist/.
  2. 2.M. Aharon, M. Elad, and A. Bruckstein. K-svd and its non-negative variant for dictionary design. In Proc. of the SPIE conference wavelets, volume 5914, 2005.
  3. 3.A. Ahmed, K. Yu, W. Xu, Y. Gong, and E. Xing. Training hierarchical feed-forward visual recognition models using transfer learning from pseudo tasks. In European Conference on Computer Vision, 2008.
  4. 4.Y. Bengio, P. Lamblin, D. Popovici, and H. Larochelle. Greedy layer-wise training of deep networks. In Advances in Neural Information Processing Systems. MIT Press, 2007.
  5. 5.A. C. Berg, T. L. Berg, and J. Malik. Shape matching and object recognition using low distortion correspondences. In CVPR, 2005.
  6. 6.N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In Proc. of Computer Vision and Pattern Recognition, 2005.
  7. 7.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 CVPR Workshop, 2004.
  8. 8.G. Hinton and R. R. Salakhutdinov. Reducing the dimensionality of data with neural networks. Science, 313(5786):504–507, 2006.
  9. 9.G. Hinton and T. Sejnowski. Unsupervised Learning: Foundations of Neural Computation. MIT press, 1999.
  10. 10.F.-J. Huang and Y. LeCun. Large-scale learning with svm and convolutional nets for generic object categorization. In Proc. Computer Vision and Pattern Recognition Conference (CVPR'06). IEEE Press, 2006.
  11. 11.K. Kavukcuoglu, M. Ranzato, R. Fergus, and Y. LeCun. Learning invariant features through topographic filter maps. In CVPR'09. IEEE, 2009.
  12. 12.K. Kavukcuoglu, M. Ranzato, and Y. LeCun. Fast inference in sparse coding algorithms with applications to object recognition. Technical report, Computational and Biological Learning Lab, Courant Institute, NYU, 2008. Tech Report CBLL-TR-2008-12-01.
  13. 13.S. Lazebnik, C. Schmid, and J. Ponce. Beyond bags of features: Spatial pyramid matching for recognizing natural scene categories. In CVPR, June 2006.
  14. 14.Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, November 1998.
  15. 15.Y. LeCun, F.-J. Huang, and L. Bottou. Learning methods for generic object recognition with invariance to pose and lighting. In Proceedings of CVPR'04. IEEE Press, 2004.
  16. 16.H. Lee, A. Battle, R. Raina, and A. Ng. Efficient sparse coding algorithms. In Advances in Neural Information Processing Systems, 2006.
  17. 17.H. Lee, E. Chaitanya, and A. Y. Ng. Sparse deep belief net model for visual area v2. In Advances in Neural Information Processing Systems, 2007.
  18. 18.H. Lee, R. Grosse, R. Ranganath, and A. Y. Ng. Convolutional deep belief networks for scalable unsupervised learning of hierarchical representations. In Proc. ICML, 2009.
  19. 19.D. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 2004.
  20. 20.S. Lyu and E. Simoncelli. Nonlinear image representation using divisive normalization. In CVPR, pages 1–8, 2008.
  21. 21.J. Mairal, F. Bach, J. Ponce, G. Sapiro, and A. Zisserman. Discriminative learned dictionaries for local image analysis. In CVPR, 2008.
  22. 22.J. Mutch and D. Lowe. Multiclass object recognition with sparse, localized features. In CVPR, 2006.
  23. 23.B. A. Olshausen and D. J. Field. Sparse coding with an overcomplete basis set: a strategy employed by v1? Vision Research, 37:3311–3325, 1997.
  24. 24.N. Pinto, D. D. Cox, and J. J. DiCarlo. Why is real-world visual object recognition hard? PLoS Computational Biology, 4(1), 2008.
  25. 25.M. Ranzato, F. Huang, Y. Boureau, and Y. LeCun. Unsupervised learning of invariant feature hierarchies with applications to object recognition. In Proc. Computer Vision and Pattern Recognition Conference (CVPR'07). IEEE Press, 2007.
  26. 26.M. Ranzato, C. Poultney, S. Chopra, and Y. LeCun. Efficient learning of sparse representations with an energy-based model. In NIPS 2006. MIT Press, 2006.
  27. 27.M. Ranzato and M. Szummer. Semi-supervised learning of compact document representations with deep networks. In Internal Conference of Machine Learning, 2008.
  28. 28.T. Serre, L. Wolf, and T. Poggio. Object recognition with features inspired by visual cortex. In CVPR, 2005.
  29. 29.M. Varma and D. Ray. Learning the discriminative power-invariance trade-off. In ICCV, 2007.
  30. 30.J. Yang, K. Yu, Y. Gong, and T. Huang. Linear spatial pyramid matching using sparse coding for image classification. In CVPR, 2009.
  31. 31.H. Zhang, A. C. Berg, M. Maire, and J. Malik. Svm-knn: Discriminative nearest neighbor classification for visual category recognition. In CVPR, 2006.

Citation

MLA
Jarrett, K., et al. “What Is the Best Multi-stage Architecture for Object Recognition?”. 2009 IEEE 12th International Conference on Computer Vision, 2009, pp. 2146–53, https://doi.org/10.1109/ICCV.2009.5459469.
APA
Jarrett, K., Kavukcuoglu, K., Ranzato, M. A., & LeCun, Y. (2009). What is the best multi-stage architecture for object recognition?. 2009 IEEE 12th International Conference on Computer Vision, 2146–2153. https://doi.org/10.1109/ICCV.2009.5459469
Chicago
Jarrett, K., K. Kavukcuoglu, M. A. Ranzato, and Y. LeCun. 2009. “What Is the Best Multi-stage Architecture for Object Recognition?”. 2009 IEEE 12th International Conference on Computer Vision, 2146–53. https://doi.org/10.1109/ICCV.2009.5459469.
Harvard
Jarrett, K. et al. (2009) “What is the best multi-stage architecture for object recognition?”, 2009 IEEE 12th International Conference on Computer Vision. IEEE, pp. 2146–2153. Available at: https://doi.org/10.1109/ICCV.2009.5459469.
Vancouver
1. Jarrett K, Kavukcuoglu K, Ranzato MA, LeCun Y (2009) What is the best multi-stage architecture for object recognition?. In: 2009 IEEE 12th International Conference on Computer Vision. IEEE, pp 2146–2153

BibTeX

@inproceedings{Jarrett_2009, title={What is the best multi-stage architecture for object recognition?}, url={http://dx.doi.org/10.1109/ICCV.2009.5459469}, DOI={10.1109/iccv.2009.5459469}, booktitle={2009 IEEE 12th International Conference on Computer Vision}, publisher={IEEE}, author={Jarrett, Kevin and Kavukcuoglu, Koray and Ranzato, Marc’ Aurelio and LeCun, Yann}, year={2009}, month=Sept, pages={2146–2153} }
Metadata:Crossref

Access the Paper

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

Open PDF

License: IEEE