Back

Example-Based Learning for View-Based Human Face Detection

K. SungT. Poggio

article1998TPAMI2,156 citations
Listen

Automated human face detection in unconstrained, complex visual scenes represents a foundational challenge for computer vision, serving as a critical prerequisite for security monitoring, automated access control, and human-computer interfaces. Traditional strategies relying on rigid correlation templates, deformable contours, or hand-engineered visual invariants often falter due to natural variations in facial expression, lighting conditions, and partial occlusions. The article addresses this operational bottleneck by developing and evaluating an example-based learning framework capable of locating upright, frontal views of human faces across multiple scales in cluttered scenes without requiring hand-crafted geometric rules.

The system standardizes visual input windows into normalized pixel patches and applies illumination corrections before comparing candidate patterns against a distribution-based reference model. This model represents image variations using several Gaussian clusters divided equally into prototypical positive face examples and carefully curated nonface examples. For each scanned image window, the system computes a difference vector representing two distinct distance components relative to each cluster centroid: a normalized Mahalanobis distance within a reduced subspace of major eigenvectors and an orthogonal Euclidean distance to that subspace. A trained multi-layer network classifier subsequently evaluates this set of distance features to decide whether a candidate image patch constitutes a human face.

The empirical evaluation demonstrated that the model achieved robust recognition accuracy across standard benchmark collections, successfully detecting approximately 96 percent of target faces in high-quality portrait datasets and roughly 80 percent on unconstrained, cluttered test images with minimal false positives. Ablation analyses confirmed that utilizing the specialized two-part distance metric consistently improved detection fidelity compared to standard Euclidean measures or full Mahalanobis formulations. Furthermore, incorporating explicit nonface clusters generated through iterative bootstrapping proved vital, substantially reducing false alarms by delineating the empirical decision boundaries separating genuine faces from naturally occurring background clutter.

These findings indicate that complex visual target detection can be effectively achieved through compact statistical density models paired with sample-efficient artificial neural classifiers, reducing the engineering overhead and fragility associated with manual rule definition. Organizations deploying automated vision systems should implement bootstrapping pipelines to capture challenging false positives during training while adopting hybrid distance metrics that balance dominant component variation with reconstruction error. However, stakeholders should note that the current implementation is restricted to vertical frontal views and shows decreased reliability when exposed to severe in-plane rotations, extreme shadows, or heavy visual degradation, indicating that future enhancements should extend the distribution modeling to wider multi-view poses and unconstrained lighting conditions.

Massachusetts Institute of TechnologyNational University of Singaporefacial imageshuman face detectionFaceHuman Face DetectionExample-Based Learning
Sung et al (1998).pdf
Cover for Example-Based Learning for View-Based Human Face Detection

Abstract

We present an example-based learning approach for locating vertical frontal views of human faces in complex scenes. The technique models the distribution of human face patterns by means of a few view-based "face" and "nonface" model clusters. At each image location, a difference feature vector is computed between the local image pattern and the distribution-based model. A trained classifier determines, based on the difference feature vector measurements, whether or not a human face exists at the current image location. We show empirically that the distance metric we adopt for computing difference feature vectors, and the "nonface" clusters we include in our distribution-based model, are both critical for the success of our system.

Topics

facial imageshuman face detection

Keywords

FaceHuman Face DetectionExample-Based LearningView-Based Human FacePATTERNMACHINE INTELLIGENCE VOLModelDistancePageIEEE TRANSACTIONS

Table of Contents

  • 1 INTRODUCTION
  • 1.1 The Face-Detection Problem
  • 1.2 Existing Face-Detection Work
  • 1.2.1 Correlation Templates
  • 1.2.2 Deformable Templates
  • 1.2.3 Image Invariants
  • 1.3 Example-Based Learning and Face Detection
  • 2 SYSTEM OVERVIEW AND APPROACH
  • 3 A DISTRIBUTION-BASED FACE MODEL
  • 3.1 Identifying and Representing the Canonical Face Manifold
  • 3.2 Preprocessing
  • 3.3 Modeling the Distribution of “Face” Patterns
  • 3.4 Modeling the Distribution of “Nonface” Patterns
  • 4 MATCHING PATTERNS WITH THE MODEL
  • 4.1 A Two-Value Distance Metric
  • 4.2 The Normalized Mahalanobis Distance
  • 4.3 The First Distance Component—Distance Within a Normalized Low-Dimensional Mahalanobis Subspace
  • 4.4 The Second Distance Component—Distance From the Low-Dimensional Mahalanobis Subspace
  • 4.5 Relationship Between Our Two-Value Distance and the Mahalanobis Distance
  • 5 THE CLASSIFIER
  • 5.1 Generating and Selecting Training Examples
  • 5.2 A Probabilistic Interpretation on Combining MLP Net Classifiers With the Distribution-Based Face Model
  • 6 RESULTS AND PERFORMANCE ANALYSIS
  • 6.1 Measuring the System’s Performance
  • 6.2 Analyzing the System’s Components
  • 6.2.1 Classifier Architecture
  • 6.2.2 The Distance Metric
  • 6.2.3 “Nonface” Model Clusters
  • 6.3 Performance Figures and Interpretation
  • 6.3.1 Classifier Architecture
  • 6.3.2 The Distance Metric
  • 6.3.3 “Nonface” Model Clusters
  • 7 CONCLUSION
  • ACKNOWLEDGMENT
  • REFERENCES

Knowls

  1. Knowl 1 — Two-Value Distance Metric for Subspace Pattern Representation

    equation

    To measure the distance between a high-dimensional image pattern and a Gaussian cluster model when sample sizes are insufficient to estimate a full-covariance matrix, the distance is decomposed into a two-value metric consisting of a normalized Mahalanobis distance within a low-dimensional principal subspace and an orthogonal Euclidean distance.

    Let x⃗∈Rd\vec{x} \in \mathbb{R}^{d} denote an unmasked image pattern vector (with d=283d = 283 unmasked pixels), and let μ⃗∈Rd\vec{\mu} \in \mathbb{R}^{d} denote the cluster centroid. Let Eh∈Rd×hE_h \in \mathbb{R}^{d \times h} be the matrix whose hh columns are the leading orthonormal eigenvectors corresponding to the h=75h = 75 largest eigenvalues of the cluster sample covariance matrix, and let Wh∈Rh×hW_h \in \mathbb{R}^{h \times h} be the diagonal matrix of these hh eigenvalues. The covariance matrix within the hh-dimensional subspace is Σh=EhWhEhT\Sigma_h = E_h W_h E_h^T.

    The first distance component D1(x⃗,μ⃗)D_1(\vec{x}, \vec{\mu}) is the normalized Mahalanobis distance between the projection of x⃗\vec{x} and the cluster centroid μ⃗\vec{\mu} in the hh-dimensional subspace:

    D1(x⃗,μ⃗)=12(hln⁡2π+ln⁡∣Σh∣+(x⃗−μ⃗)TΣh−1(x⃗−μ⃗))D_1(\vec{x}, \vec{\mu}) = \frac{1}{2}\left( h \ln 2\pi + \ln |\Sigma_h| + (\vec{x} - \vec{\mu})^T \Sigma_h^{-1} (\vec{x} - \vec{\mu}) \right)

    where ∣Σh∣=∏i=1hλi|\Sigma_h| = \prod_{i=1}^h \lambda_i is the determinant of the subspace covariance (the product of the hh largest eigenvalues), and Σh−1=EhWh−1EhT\Sigma_h^{-1} = E_h W_h^{-1} E_h^T.

    The second distance component D2(x⃗,μ⃗)D_2(\vec{x}, \vec{\mu}) is the Euclidean distance between x⃗\vec{x} and its orthogonal projection x⃗p=μ⃗+EhEhT(x⃗−μ⃗)\vec{x}_p = \vec{\mu} + E_h E_h^T (\vec{x} - \vec{\mu}) onto the hh-dimensional subspace:

    D2(x⃗,μ⃗)=∥x⃗−x⃗p∥=∥(I−EhEhT)(x⃗−μ⃗)∥D_2(\vec{x}, \vec{\mu}) = \|\vec{x} - \vec{x}_p\| = \|(I - E_h E_h^T)(\vec{x} - \vec{\mu})\|

    Component D1D_1 captures pattern variations along the dominant directions of cluster variance where empirical statistics are reliable, while component D2D_2 assumes isotropic Gaussian noise across the remaining d−hd - h dimensions where empirical eigenvalue estimates are unreliable.

  2. Knowl 2 — Distribution-Based Face Detection System Pipeline

    model/method

    The view-based face detection system identifies vertical frontal human faces across unknown positions and scales in grayscale images using a multi-stage detection pipeline:

    1. Multi-Scale Scanning: The input image is searched exhaustively using a 19×1919 \times 19 pixel sliding window shifted pixel-by-pixel. To detect faces at larger scales, the image is iteratively downscaled by a scale factor of 1.21.2 per octave step, and scanned with the same 19×1919 \times 19 window up to maximum face sizes of 100×100100 \times 100 pixels.

    2. Window Preprocessing: Each candidate 19×1919 \times 19 window pattern is normalized through circular masking (reducing dimensionality from 361 to 283 pixels), linear brightness plane subtraction, and histogram equalization.

    3. Difference Feature Vector Computation: The preprocessed pattern is compared against a canonical distribution-based model consisting of 12 multi-dimensional Gaussian clusters (6 face clusters and 6 non-face clusters). For each cluster, a two-value distance (D1,D2)(D_1, D_2) is computed, resulting in a 24-dimensional distance feature vector.

    4. Classifier Evaluation: A trained Multilayer Perceptron (MLP) receives the 24-dimensional feature vector and classifies the local window as either a face (output near 1) or non-face (output near 0). Detections across multiple overlapping windows and scales are grouped to locate each face.

  3. Knowl 3 — Elliptical k-Means Clustering Algorithm

    algorithm

    The elliptical kk-means clustering algorithm partitions high-dimensional pattern vectors into kk directionally elongated Gaussian clusters by estimating both cluster centroids and cluster covariance matrices using a normalized Mahalanobis distance metric.

    Input: Data samples X={x⃗1,…,x⃗N}⊂RdX = \{\vec{x}_1, \dots, \vec{x}_N\} \subset \mathbb{R}^d, number of clusters kk, maximum inner iterations ImaxI_{max}, maximum outer iterations OmaxO_{max}
    Output: Cluster centroids {μ⃗1,…,μ⃗k}\{\vec{\mu}_1, \dots, \vec{\mu}_k\} and covariance matrices {Σ1,…,Σk}\{\Sigma_1, \dots, \Sigma_k\}
    Initialize cluster centroids {μ⃗1,…,μ⃗k}\{\vec{\mu}_1, \dots, \vec{\mu}_k\} using standard Euclidean vector quantization on XX
    Initialize covariance matrices Σj←Id×d\Sigma_j \leftarrow I_{d \times d} for all j∈{1,…,k}j \in \{1, \dots, k\}
    outer_iter ←0\leftarrow 0
    repeat
        inner_iter ←0\leftarrow 0
        repeat
            Assign each sample x⃗i∈X\vec{x}_i \in X to the nearest cluster jj minimizing the normalized Mahalanobis distance Md(x⃗i,μ⃗j)=12(dln⁡2π+ln⁡∣Σj∣+(x⃗i−μ⃗j)TΣj−1(x⃗i−μ⃗j))M_d(\vec{x}_i, \vec{\mu}_j) = \frac{1}{2}(d \ln 2\pi + \ln |\Sigma_j| + (\vec{x}_i - \vec{\mu}_j)^T \Sigma_j^{-1} (\vec{x}_i - \vec{\mu}_j))
            Recompute cluster centroids μ⃗j←1∣Sj∣∑x⃗∈Sjx⃗\vec{\mu}_j \leftarrow \frac{1}{|S_j|} \sum_{\vec{x} \in S_j} \vec{x} for all partitions SjS_j
            inner_iter ←\leftarrow inner_iter +1+ 1
        until sample partition assignments remain unchanged or inner_iter ≥Imax\ge I_{max}
        Recompute covariance matrices Σj←1∣Sj∣∑x⃗∈Sj(x⃗−μ⃗j)(x⃗−μ⃗j)T\Sigma_j \leftarrow \frac{1}{|S_j|} \sum_{\vec{x} \in S_j} (\vec{x} - \vec{\mu}_j)(\vec{x} - \vec{\mu}_j)^T for all j∈{1,…,k}j \in \{1, \dots, k\}
        Assign each sample x⃗i∈X\vec{x}_i \in X to the nearest cluster jj using the updated normalized Mahalanobis distance
        outer_iter ←\leftarrow outer_iter +1+ 1
    until sample partition assignments remain unchanged or outer_iter ≥Omax\ge O_{max}
    return {μ⃗1,…,μ⃗k}\{\vec{\mu}_1, \dots, \vec{\mu}_k\} and {Σ1,…,Σk}\{\Sigma_1, \dots, \Sigma_k\}

  4. Knowl 4 — Bootstrap Strategy for Hard Negative Example Mining

    algorithm

    Because the space of non-face images is virtually infinite, an active learning bootstrap procedure incrementally discovers and selects high-utility "face-like" negative examples to define the boundary of the face distribution manifold.

    Input: Initial positive face database DfaceD_{face}, initial small negative non-face database DnonfaceD_{nonface}, sequence of images containing no human faces InegI_{neg}
    Output: Trained classifier and enriched non-face database DnonfaceD_{nonface}
    repeat
        Train the face classifier using the current training databases DfaceD_{face} and DnonfaceD_{nonface}
        Run the trained face detector across all scales and locations over the non-face image sequence InegI_{neg}
        Collect all false positive windows FPFP detected by the system (non-face patterns misclassified as faces)
        Add the collected false positive patterns to the negative training database: Dnonface←Dnonface∪FPD_{nonface} \leftarrow D_{nonface} \cup FP
    until false positive rate on validation images falls below the desired operating threshold
    return trained classifier and DnonfaceD_{nonface}

    This process collected 6,189 face-like negative patterns starting from a small initial set, ensuring the negative training set concentrates on non-face patterns situated nearest to the face distribution manifold.

  5. Knowl 5 — Four-Stage Image Preprocessing Pipeline

    model/method

    To reduce extraneous photometric and geometric variations, every 19×1919 \times 19 candidate image window is processed through four deterministic normalization steps before distance computation:

    1. Window Resizing: The input patch is resized to a standardized 19×1919 \times 19 pixel grid, keeping dimensionality manageable while preserving key facial structures (eyes, nose, mouth).

    2. Masking: A predefined circular 19×1919 \times 19 binary mask is applied to remove near-boundary corner pixels corresponding to hair and background clutter. This reduces the effective input vector dimensionality from 361 to d=283d = 283 pixels.

    3. Illumination Gradient Correction: A best-fit linear brightness plane I(u,v)=au+bv+cI(u, v) = a u + b v + c is fitted to the unmasked pixel intensities using linear least squares and subtracted from the window. This eliminates linear shading gradients and reduces heavy shadows caused by extreme side lighting.

    4. Histogram Equalization: The brightness histogram of the unmasked pixels is equalized to normalize the global image contrast across varied illumination levels and camera sensor response curves.

  6. Knowl 6 — Gaussian Mixture Modeling of Face and Non-Face Manifolds

    model/method

    The distribution of frontal face views and face-like patterns in the 283-dimensional masked pixel space is modeled using two sets of multidimensional Gaussian clusters:

    • Face Model Clusters: An empirical distribution of 4,150 normalized frontal face patterns (synthesized from 1,067 real hand-cropped face images augmented by horizontal mirroring and small planar rotations of ±5∘\pm 5^\circ) is modeled with 6 Gaussian clusters via elliptical kk-means clustering.
    • Non-Face Model Clusters: An empirical distribution of 6,189 face-like negative patterns (obtained via iterative bootstrap mining) is modeled with 6 Gaussian clusters.

    Each of the 12 clusters is defined by a 283-dimensional centroid μ⃗k\vec{\mu}_k and a local covariance matrix parameterized by its 75 leading orthonormal eigenvectors and eigenvalues. The 6 non-face clusters act as negative anchor prototypes, carving out boundaries and non-face sub-regions adjacent to the canonical face manifold.

  7. Knowl 7 — Multilayer Perceptron Architecture for Distance-Based Face Classification

    model/method

    Classification of candidate image windows is performed by a Multi-Layer Perceptron (MLP) trained on distance feature vectors:

    • Input Layer: 24 input terminals receiving 12 pairs of (D1,D2)(D_1, D_2) distance values, representing the Mahalanobis subspace distance (D1D_1) and orthogonal Euclidean distance (D2D_2) from the candidate pattern to each of the 12 Gaussian cluster centroids (6 face centroids and 6 non-face centroids).
    • Hidden Layer: A single hidden layer consisting of 24 hidden units with sigmoidal activation functions σ(z)=11+e−z\sigma(z) = \frac{1}{1 + e^{-z}}.
    • Output Layer: A single output unit with sigmoidal activation producing a score thresholded to classify the input as face (11) or non-face (00).
    • Training: Trained using standard error back-propagation on a dataset of 47,316 pattern vectors (4,150 positive face examples and 43,166 negative bootstrap-mined non-face examples) until output mean squared error stabilizes.

  8. Knowl 8 — Face Detection Performance on Real-World Benchmark Databases

    empirical result

    The face detection system was evaluated on two test datasets across varying image quality, lighting, and background complexity:

    • Database 1 (High Quality / Best Case): Consisting of 301 frontal and near-frontal face mugshots of 71 distinct individuals with moderate lighting variations, the system achieved a 96.3% detection rate (correctly detecting 290 out of 301 faces) with 3 false positives. The few missed faces had strong illumination shadows or out-of-plane head rotations.
    • Database 2 (Complex Scenes / Average Case): Consisting of 23 complex images containing 149 faces with wide quality variations (including CCD camera images and low-quality newspaper scans with complex backgrounds), the system achieved a 79.9% detection rate (119 out of 149 faces) with 5 false positives. Missed faces were predominantly low-quality half-tone newspaper scans, hand-drawn faces, or faces with severe off-plane rotations.

  9. Knowl 9 — Detection Rate and False Alarm Comparison Across Distance Metrics and Classifier Architectures

    data/table

    Performance across different distance metrics and classifier configurations was evaluated on Test Database 1 (301 faces) and Test Database 2 (149 faces). The two-value distance metric consistently provided superior detection rates and lower false alarms compared to using either distance component alone or the standard full-space Mahalanobis distance.

    Classifier 2-Value Component D1D_1 Component D2D_2 Std. Mahalanobis
    Architecture DB1 DB2 DB1 DB2 DB1 DB2 DB1 DB2
    Multi-layer 96.3% (3) 79.9% (5) 91.6% (21) 85.1% (114) 91.4% (4) 65.1% (5) 84.1% (9) 42.6% (5)
    Perceptron
    Single Unit 96.7% (3) 84.6% (13) 93.3% (15) 85.1% (94) 92.3% (3) 68.2% (5) 93.0% (13) 58.6% (11)
    Perceptron
    Nearest 65.1% (1) – 97.4% (208) – 53.9% (1) – 71.8% (5) –
    Neighbor

    Format for entries: detection rate percentage followed by total number of false positive detections in parentheses. Nearest neighbor yields poor trade-offs on Database 1 due to extreme sample imbalance (43,166 non-face samples vs. 4,150 face samples), which causes query boundary points to be dominated by non-face neighbors. Both network-based classifiers perform significantly better when supplied with the full two-value distance feature vector.

  10. Knowl 10 — Impact of Non-Face Prototype Clusters on Detection and False Positive Rates

    data/table

    An ablation study evaluated the contribution of negative ("non-face") prototype clusters in the distribution-based model across Test Database 1 (301 faces) and Test Database 2 (149 faces).

    Classifier 6 Face 6 Non-Face 12 Face Clusters 6 Face Clusters
    Architecture DB1 DB2 DB1 DB2 DB1 DB2
    Multi-layer Perceptron 96.3% (3) 79.9% (5) 85.3% (21) 69.6% (74) 59.7% (17) 60.9% (41)
    Single Perceptron 96.7% (3) 84.6% (13) 52.1% (6) 49.7% (16) 66.6% (25) 55.4% (56)

    Format for entries: detection rate percentage followed by total number of false positive detections in parentheses. Incorporating 6 non-face clusters alongside 6 face clusters significantly improves detection rates while suppressing false alarms compared to using 12 face-only clusters or 6 face-only clusters. The non-face clusters provide explicit reference points that carve out negative distribution regions near the face manifold boundary.

  11. Knowl 11 — Limitations Regarding Out-of-Plane Pose Variations and Dark Distractor Regions

    limitation

    The example-based face detection system has several specific operational limitations:

    1. Pose Sensitivity: The model is trained exclusively on vertical frontal and near-frontal views (±5∘ \pm 5^\circ in-plane rotation); it fails to detect faces with large out-of-plane rotation (profile or semi-profile views) or significant in-plane tilt.
    2. Extreme Illumination Shadows: Despite linear brightness gradient subtraction and histogram equalization, severe directional lighting that casts deep shadows over critical subfeatures (such as one half of the face or the eye sockets) causes missed detections.
    3. False Positives from Eye-Like Features: Most false positive detections generated by the system occur on background texture patches containing dark paired spots or horizontal dark bands that mimic the spatial luminance structure of human eyes.

Coverage note — None was omitted; all primary contributions, models, algorithms, experimental evaluations, tabular ablations, and stated limitations have been fully captured.

References

  1. 1.D. Beymer, A. Shashua, and T. Poggio, "Example Based Image Analysis and Synthesis," AIM-1431, Artificial Intelligence Laboratory, Massachusetts Institute of Technology, 1993.
  2. 2.M. Bichsel, "Strategies of Robust Objects Recognition for Automatic Identification of Human Faces," PhD thesis, ETH, Zurich, 1991.
  3. 3.R. Brunelli and T. Poggio, "Face Recognition: Features Versus Templates," IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 15, no. 10, pp. 1,042–1,052, Oct. 1993.
  4. 4.M.C. Burl, U. Fayyad, P. Perona, P. Smyth, and M.P. Burl, "A Trainable Tool for Finding Small Volcanoes in SAR Imagery of Venus," CNS TR-34, Calif. Inst. of Technology, 1993.
  5. 5.R.O. Duda and P.E. Hart, Pattern Classification and Scene Analysis. New York: John Wiley and Sons Inc., 1973.
  6. 6.W.E.L. Grimson and T. Lozano-Perez, "Model-Based Recognition and Localization From Sparse Range Data," A. Rosenfeld, ed., Techniques for 3-D Machine Perception. Amsterdam: North-Holland, 1985.
  7. 7.G. Hinton, M. Revow, and P. Dayan, "Recognizing Handwritten Digits Using Mixture of Linear Models," Proc. Advances in Neural Information Processings Systems, vol. 7, 1995.
  8. 8.M. Kirby and L. Sirovich, "Applications of the Karhunen-Loeve Procedure for the Characterization of Human Faces," IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 12, no. 1, pp. 103–108, Jan. 1990.
  9. 9.B. Kumar, D. Casasent, and H. Murakami, "Principal Component Imagery for Statistical Pattern Recognition Correlators," Optical Eng., vol. 21, no. 1, Jan./Feb. 1982.
  10. 10.A. Mahalanobis, A. Forman, N. Day, M. Bower, and R. Cherry, "Multi-Class SAR ATR Using Shift-Invariant Correlation Filters," Pattern Recognition, vol. 27, no. 4, pp. 619–626, 1994.
  11. 11.M. Oren, C. Papageorgiou, P. Sinha, E. Osuna, and T. Poggio, "Pedestrian Detection Using Wavelet Templates," Proc. CVPR, 1997.
  12. 12.A. Pentland, B. Moghaddam, and T. Starner, "View-Based and Modular Eigenspaces for Face Recognition," Proc. CVPR, pp. 84–91, 1994.
  13. 13.T. Poggio and T. Vetter, "Recognition and Structure From One (2D) Model View: Observations on Prototypes, Object Classes, and Symmetries," AIM-1347, Artificial Intelligence Laboratory, Massachusetts Institute of Technology, 1992.
  14. 14.D. Rumelhart and J. McClelland, Parallel Distributed Processing, vol. 1. Cambridge, Mass.: MIT Press, 1986.
  15. 15.P. Sinha, "Object Recognition via Image Invariants: A Case Study," Investigative Ophthalmology and Visual Science, vol. 35, pp. 1,735–1,740, May 1994.
  16. 16.K. Sung and T. Poggio, "Example-Based Learning for View-Based Human Face Detection," Proc. Image Understanding Workshop, vol. 2, pp. 843–850, 1994.
  17. 17.K. Sung, "Learning and Example Selection for Object and Pattern Detection," PhD thesis, Massachusetts Institute of Technology, 1995.
  18. 18.M. Turk and A. Pentland, "Eigenfaces for Recognition," J. Cognitive Neuroscience, vol. 3, no. 1, pp. 71–86, 1991.
  19. 19.A. Yuille, P. Hallinan, and D. Cohen, "Feature Extraction From Faces Using Deformable Templates," Int'l J. Computer Vision, vol. 8, no. 2, pp. 99–111, 1992.

Citation

MLA
Sung, K.-K., and T. Poggio. “Example-based Learning for View-based Human Face Detection”. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 20, no. 1, 1998, pp. 39–51, https://doi.org/10.1109/34.655648.
APA
Sung, K.-K., & Poggio, T. (1998). Example-based learning for view-based human face detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(1), 39–51. https://doi.org/10.1109/34.655648
Chicago
Sung, K.-K., and T. Poggio. 1998. “Example-based Learning for View-based Human Face Detection”. IEEE Transactions on Pattern Analysis and Machine Intelligence 20 (1): 39–51. https://doi.org/10.1109/34.655648.
Harvard
Sung, K.-K. and Poggio, T. (1998) “Example-based learning for view-based human face detection”, IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(1), pp. 39–51. Available at: https://doi.org/10.1109/34.655648.
Vancouver
1. Sung K-K, Poggio T (1998) Example-based learning for view-based human face detection. IEEE Transactions on Pattern Analysis and Machine Intelligence 20:39–51

BibTeX

@article{Sung_1998, title={Example-based learning for view-based human face detection}, volume={20}, ISSN={0162-8828}, url={http://dx.doi.org/10.1109/34.655648}, DOI={10.1109/34.655648}, number={1}, journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, publisher={Institute of Electrical and Electronics Engineers (IEEE)}, author={Sung, K.-K. and Poggio, T.}, year={1998}, pages={39–51} }
Metadata:Crossref

Access the Paper

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

Open PDF