Attribute and simile classifiers for face verification

Neeraj KumarA. BergP. BelhumeurS. Nayar

article2009ICCV1,721 citations

Proposes novel attribute and simile classifiers that capture high-level visual traits and reference similarities, dramatically cutting face verification error rates on unconstrained benchmarks without requiring image pair alignment.

Listen

Automatic face verification systems often struggle in unconstrained, real-world conditions where lighting, pose, facial expressions, and camera quality vary widely. While conventional methods frequently fail under these uncontrolled settings or rely on computationally expensive and brittle image-alignment processes, humans can verify identities across varied conditions with exceptional accuracy.

The article evaluates two high-level visual trait methods—attribute classifiers and simile classifiers—to determine whether extracting human-interpretable and reference-based visual characteristics can improve unconstrained face verification without requiring pairwise image alignment.

The researchers developed two distinct classification approaches. The attribute classifier trains binary support vector machines on 65 describable visual characteristics such as gender, race, age, and facial features, utilizing over 125,000 crowd-sourced image labels for training. The simile classifier removes the need for manual labeling by training binary classifiers to measure how closely specific regions of an unaligned face resemble corresponding regions across 60 reference individuals. Both techniques generate compact visual trait vectors for each image independently. The system then compares these trait vectors using a separate verification classifier. The models were benchmarked on the standard Labeled Faces in the Wild dataset and a newly compiled dataset of public figures called PubFig, which comprises 60,000 web-collected images across 200 individuals.

The study established several key findings. First, both trait-based approaches significantly outperform previous state-of-the-art benchmarks on Labeled Faces in the Wild: attribute classifiers reduced error rates by 23.92%, simile classifiers reduced error rates by 26.34%, and a hybrid combination of both reduced error rates by 31.68%, achieving an overall accuracy of 85.29%. Second, crowd-sourced baseline testing revealed human verification accuracy on the same benchmark reaches 99.20% on full images and 97.53% on tightly cropped faces. Third, human testing showed an accuracy of 94.27% when the face was entirely obscured and only background and context were visible, demonstrating that background context can easily bias evaluations if not strictly masked out during algorithm training.

These results indicate that high-level visual trait representations offer a robust, computationally efficient alternative to traditional low-level pixel alignment. By computing compact descriptors for each image independently, systems can scale more effectively without costly pairwise matching. Furthermore, the ability to train simile classifiers without manual annotation provides a scalable path to build robust visual models at lower operational costs. However, the large gap between the hybrid algorithm's 85.29% accuracy and human performance of 97.53% on cropped faces indicates substantial room for progress before automated systems can match human reliability in uncontrolled environments.

Decision-makers and practitioners deploying face verification in real-world scenarios should adopt high-level visual trait frameworks and strictly enforce facial masking to prevent background context from artificially inflating accuracy. Organizations should leverage the newly released, deeper PubFig dataset to evaluate algorithmic resilience against explicit variations in lighting, pose, and expression. Future development should focus on expanding the reference pool for simile classifiers and scaling recognition experiments across more diverse identity sets.

A primary operational limitation is the initial data acquisition requirement: attribute classifiers demand large volumes of high-quality manual labels, whereas simile classifiers depend on having multiple diverse images per reference individual. Additionally, automated face and fiducial detection systems can occasionally fail to detect faces in highly unconstrained photos, which slightly impacts automated pipeline consistency. Overall, the methodology provides strong, statistically validated improvements over prior methods on standard benchmarks, but caution is warranted when deploying in production environments where non-frontal poses and extreme lighting remain challenging.

Columbia University asian continental ancestry group automated systems binary classification face verification attribute classifiers simile classifiers

Kumar et al (2009).pdf

Paper: Describing Objects by their Attributes, Ali Farhadi et al. (2009). This paper establishes the foundational computer vision paradigm of using semantic attribute classifiers for descriptive visual recognition that directly precedes high-level facial trait classification.
Paper: Learning a similarity metric discriminatively, with application to face verification, Sumit Chopra et al. (2005). It introduces discriminative metric learning and Siamese networks for unconstrained face verification, defining the pair comparison problem addressed by attribute and simile classifiers.
Paper: Eigenfaces vs. Fisherfaces: Recognition Using Class Specific Linear Projection, Peter N. Belhumeur et al. (1996). This classic work details low-level projection baselines and the fundamental challenges that lighting and expression variations pose to traditional face verification.
Paper: Face Recognition Based on Fitting a 3D Morphable Model, Volker Blanz et al. (2003). It presents 3D morphable model fitting as an explicit alignment strategy to handle pose and lighting variations, serving as a primary point of comparison for alignment-free trait methods.
Paper: The CMU Pose, Illumination, and Expression Database, Terence Sim et al. (2003). This paper introduces the CMU PIE benchmark systematically isolating pose, illumination, and expression variations in facial recognition.
Paper: Active Appearance Models Revisited, Iain Matthews et al. (2004). It provides the standard formulation for Active Appearance Models, representing the generative alignment and fiducial fitting techniques that trait classifiers seek to avoid or streamline.
Paper: Detecting Faces in Images: A Survey, Ming-Hsuan Yang et al. (2002). This survey provides a comprehensive analysis of early appearance-based and feature-based face detection and localization pipelines.

Paper: Deep Learning Face Attributes in the Wild, Ziwei Liu et al. (2015). It advances the facial attribute classification paradigm by replacing shallow classifiers with deep convolutional networks trained directly on unconstrained in-the-wild facial attributes.
Paper: Deep Learning Face Representation by Joint Identification-Verification, Yi Sun et al. (2014). This work evolves face verification from hand-crafted attribute descriptors to end-to-end deep representation learning using joint identification and verification supervision.
Paper: Learning Face Representation from Scratch, Dong Yi et al. (2014). It builds upon the public dataset and benchmark evaluation needs highlighted by PubFig by introducing the CASIA-WebFace dataset for training modern deep facial representations.
Paper: Deep Face Recognition, Omkar M. Parkhi et al. (2015). It demonstrates how large-scale curated web datasets and deep metric learning supersede intermediate attribute-based representations for unconstrained face verification.
Paper: Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification, Joy Buolamwini et al. (2018). It directly critiques and evaluates subgroup disparities and demographic biases in commercial facial attribute and gender classification pipelines.
Paper: ArcFace: Additive Angular Margin Loss for Deep Face Recognition, Jiankang Deng et al. (2018). It represents the modern culmination of unconstrained face verification by utilizing additive angular margin losses to learn robust hyperspherical feature embeddings.
Paper: MS-Celeb-1M: A Dataset and Benchmark for Large-Scale Face Recognition, Yandong Guo et al. (2016). It scales the public figure dataset and entity disambiguation concept introduced with PubFig up to a million identities.
Paper: VGGFace2: A Dataset for Recognising Faces across Pose and Age, Qiong Cao et al. (2017). It provides a massive-scale dataset explicitly designed to model variations in pose and age across thousands of identities for deep facial recognition.

Cover for Attribute and simile classifiers for face verification

Abstract

We present two novel methods for face verification. Our first method – “attribute” classifiers – uses binary classifiers trained to recognize the presence or absence of describable aspects of visual appearance (e.g., gender, race, and age). Our second method – “simile” classifiers – removes the manual labeling required for attribute classification and instead learns the similarity of faces, or regions of faces, to specific reference people. Neither method requires costly, often brittle, alignment between image pairs; yet, both methods produce compact visual descriptions, and work on real-world images. Furthermore, both the attribute and simile classifiers improve on the current state-of-the-art for the LFW data set, reducing the error rates compared to the current best by 23.92% and 26.34%, respectively, and 31.68% when combined. For further testing across pose, illumination, and expression, we introduce a new data set – termed PubFig – of real-world images of public figures (celebrities and politicians) acquired from the internet. This data set is both larger (60,000 images) and deeper (300 images per individual) than existing data sets of its kind. Finally, we present an evaluation of human performance.

Topics

asian continental ancestry group automated systems binary classification classification models descriptors facial expression facial images hybrid algorithms support vector verification method

Keywords

face verification attribute classifiers simile classifiers describable attributes face recognition trait classifiers LFW benchmark support vector machines

1. Introduction
2. Related Work
3. Our Approach
3.1. Low-level Features
3.2. Attribute Classifiers
3.3. Simile Classifiers
3.4. Verification Classifier
4. Experiments
4.1. Labeled Faces in the Wild
4.2. Human Performance on LFW
4.3. PubFig Data Set
5. Discussion
References

Knowls

Knowl 1 — Trait-Based Face Verification Decision Pipeline
model/method
Face verification between two unaligned face images $I_1$ and $I_2$ is performed by comparing high-level visual trait vectors extracted independently from each image without requiring pairwise spatial alignment.

For an input face image $I$ , a collection of $k$ low-level feature descriptors $f_{i=1 \dots k}(I)$ are extracted and concatenated into a low-level feature vector $F(I) = \langle f_1(I), \dots, f_k(I) \rangle$ . A set of $n$ trained binary trait classifiers $C_{i=1 \dots n}$ (which can be attribute classifiers, simile classifiers, or a concatenation of both) is evaluated on $F(I)$ to generate an $n$ -dimensional continuous trait score vector:

$C(I) = \langle C_1(F(I)), \dots, C_n(F(I)) \rangle$

where each classifier output lies in the range $[-1, 1]$ . Let $a_i = C_i(F(I_1))$ and $b_i = C_i(F(I_2))$ denote the respective trait outputs for trait $i \in \{1, \dots, n\}$ . For each trait $i$ , a two-dimensional comparison pair $p_i$ is constructed:

$p_i = \left(|a_i - b_i|,\, (a_i \cdot b_i)\right) \cdot g\left(\frac{1}{2}(a_i + b_i)\right)$

where $g(x) = \frac{1}{\sqrt{2\pi}} \exp\left(-\frac{x^2}{2}\right)$ is the probability density function of a standard normal distribution (mean zero, variance one). The first term measures absolute score discrepancy, the second term enforces sign consistency (rewarding agreements where both faces have matching positive or matching negative traits), and the Gaussian weighting assigns higher importance to differences near the decision boundary (zero) where SVM margins operate.

The vector $\langle p_1, \dots, p_n \rangle \in \mathbb{R}^{2n}$ (optionally augmented with 6 metadata features: three pose angles, pose confidence, and two image/file size quality measures) is passed into a final verification classifier $D$ parameterized as a Support Vector Machine (SVM) with a Radial Basis Function (RBF) kernel:

$v(I_1, I_2) = D(p_1, \dots, p_n)$

The sign of $v(I_1, I_2)$ determines the verification decision: $v(I_1, I_2) > 0$ indicates that $I_1$ and $I_2$ depict the same person, while $v(I_1, I_2) \le 0$ indicates different identities.
Knowl 2 — Reference-Based Simile Classifiers for Facial Representation
model/method
Simile classifiers represent faces by measuring the similarity of specific localized facial regions to a fixed set of reference individuals, eliminating the need for manual semantic annotation of facial attributes.

To construct simile classifiers:
1. A gallery of reference individuals (60 individuals) is collected, where all reference individuals are completely disjoint in identity from the evaluation and test sets.
2. For each reference person, up to 600 positive face images of that person across varied unconstrained conditions are paired with negative examples randomly sampled from other individuals (up to 10 times the number of positive examples).
3. Classifiers are trained for 8 localized facial regions (e.g., eyebrows, eyes, nose, mouth) and 6 low-level feature types extracted from the canonical rectified face representation.
4. For each region and feature combination of a reference person, an SVM with an RBF kernel is trained via greedy forward feature selection to discriminate that person's facial subregion from the general population.
When evaluated on an arbitrary unseen face image, the simile classifiers output continuous similarity scores indicating how closely each region of the unseen face resembles the corresponding region of the reference individuals across multiple views rather than matching against a single template.
Knowl 3 — Semantic Attribute Classifiers for Facial Representation
model/method
Attribute classifiers are binary classifiers trained to detect 65 human-describable semantic visual traits on faces, encompassing demographic categories (e.g., gender, race, age bracket), facial hair (e.g., mustache, goatee, no beard), facial structure (e.g., round face, nose shape, high cheekbones), expressions (e.g., smiling, frowning), accessories (e.g., eyeglasses, sunglasses, lipstick), and imaging conditions (e.g., flash lighting, blurry, soft lighting).

Training is conducted as follows:
1. For each attribute, at least 1,000 labeled face images (at least 500 positive and 500 negative) are acquired via crowdsourced human annotations where 3 independent annotators must reach unanimous agreement.
2. A pool of candidate low-level features is extracted across hand-labeled, pose-enlarged facial regions.
3. A greedy forward feature selection algorithm iteratively selects up to 6 low-level features per attribute, choosing at each step the candidate feature whose addition to the current feature set yields the greatest reduction in training error.
4. Using the selected low-level features, an SVM with a Radial Basis Function (RBF) kernel is trained to output a decision score in the range $[-1, 1]$ reflecting the presence or absence of the attribute.
Knowl 4 — Face Verification Performance on the Labeled Faces in the Wild Benchmark
empirical result
Evaluated on the Labeled Faces in the Wild (LFW) benchmark under the standard View 2 image-restricted 10-fold cross-validation protocol (6,000 image pairs across 5,749 individuals), trait-based face verification achieves superior accuracy over prior state-of-the-art approaches:
- Attribute Classifiers (65 traits): $83.62\% \pm 1.58\%$ accuracy (a $23.92\%$ relative error rate reduction compared to the prior best method).
- Simile Classifiers (60 reference people): $84.14\% \pm 1.31\%$ accuracy (a $26.34\%$ relative error rate reduction compared to the prior best method).
- Hybrid Classifiers (Attributes + Similes combined): $85.29\% \pm 1.23\%$ accuracy (a $31.68\%$ relative error rate reduction compared to the prior best method).
For comparison, previous methods evaluated on LFW reported lower verification accuracies: Wolf et al. descriptor-based funneled method achieved $78.47\%$ , Merl+Nowak funneled achieved $76.18\%$ , and Nowak funneled achieved $73.93\%$ . All trait classifiers operate solely on face regions with background masked out and without requiring pairwise spatial registration between test image pairs.
Knowl 5 — Human Verification Accuracy and Background Context Bias on LFW
empirical result
Human face verification performance on the 6,000 evaluation pairs of the Labeled Faces in the Wild (LFW) dataset was measured across three visual presentation conditions using crowdsourced annotations (10 independent human judgements per pair, totaling 240,000 decisions with confidence ratings):
1. Original Images (full unconstrained photo with hair, neck, and environment): Human verification accuracy reached $99.20\%$ .
2. Tight Face Crop (background and contextual regions blacked out, showing only the internal face including eyes, nose, mouth, and minimal hair/ears): Human verification accuracy dropped to $97.53\%$ , representing more than a threefold increase in human error rate compared to original images.
3. Inverse Crop (internal face blacked out, displaying exclusively the background, neck, torso, and external context): Humans achieved $94.27\%$ verification accuracy.
This demonstrates that uncropped real-world face datasets provide strong non-facial context and background cues that can artificially inflate verification performance if algorithms do not mask out regions outside the face.
Knowl 6 — Low-Level Facial Feature Extraction and Normalization Pipeline
model/method
To extract low-level facial features resistant to unconstrained imaging variations:
1. Face Detection and Alignment: A commercial face detector detects faces and fiducial landmarks. An affine warp maps fiducials to a canonical coordinate frame.
2. Regional Partitioning: Rectified face images are partitioned into enlarged hand-labeled subregions corresponding to specific facial components (e.g., eyes, nose, mouth, eyebrows, forehead, jaw).
3. Feature Types: Within each region, visual representations are extracted across multiple modalities: raw pixel intensities in RGB and HSV color spaces, edge magnitudes, and image gradient directions.
4. Normalization: Extracted values undergo one of three normalization schemes: standard score normalization (subtracting the regional mean and dividing by standard deviation), mean normalization (dividing by the regional mean), or no normalization.
5. Aggregation: Normalized values are pooled into final low-level descriptor vectors $f_i$ via direct spatial concatenation, histogram binning, or statistical moments (mean and variance).
Knowl 7 — Public Figures (PubFig) Dataset and Benchmark Protocol
experimental setup
The Public Figures (PubFig) dataset comprises 60,000 real-world unconstrained internet images spanning 200 public figures (celebrities and politicians), averaging approximately 300 images per subject.

Two verification benchmarks are established on PubFig:
1. 20,000-Pair Verification Benchmark: 20,000 pairs across 140 individuals (disjoint from the 60 reference people used for simile training), organized into 10 cross-validation folds with mutually disjoint sets of 14 individuals per fold.
2. Confounding Factor Stratified Subsets: Test pairs from the 20,000-pair benchmark are partitioned into paired "easy" and "difficult" evaluation subsets across three attributes:
  - Pose: "Easy" pairs have both faces frontal (pitch and yaw both $< 10^\circ$ ); "difficult" pairs contain at least one non-frontal face.
  - Illumination: "Easy" pairs have both images frontally lit; "difficult" pairs contain non-frontal illumination.
  - Expression: "Easy" pairs have both images exhibiting neutral expressions; "difficult" pairs contain at least one non-neutral expression (e.g., smiling, talking, frowning).
Knowl 8 — PubFig Verification Performance across Pose, Illumination, and Expression Variations
empirical result
Face verification using attribute classifiers on the PubFig benchmark yields an overall accuracy of $77.78\%$ across the full 20,000-pair evaluation set. Evaluating on subsets stratified by confounding imaging factors reveals differential sensitivity:
- Pose: "Easy" (frontal pose $< 10^\circ$ pitch/yaw) achieves $80.81\%$ accuracy; "Difficult" (non-frontal pose) achieves $77.50\%$ accuracy (a performance drop of $3.31$ percentage points).
- Illumination: "Easy" (frontal lighting) achieves $79.54\%$ accuracy; "Difficult" (non-frontal lighting) achieves $75.32\%$ accuracy (a performance drop of $4.22$ percentage points).
- Expression: "Easy" (neutral expressions) achieves $78.40\%$ accuracy; "Difficult" (non-neutral expressions) achieves $77.66\%$ accuracy (a performance drop of $0.74$ percentage points).
Among the three confounding factors, illumination variation causes the largest degradation in verification accuracy, followed by pose variation, whereas expression variation causes the smallest relative degradation.

Knowl 9 — Accuracies of 65 Semantic Facial Attribute Classifiers

data/table

The table below lists the individual classification accuracies on held-out face images across all 65 binary attribute classifiers trained via greedy forward feature selection and RBF SVMs on non-frontal and unconstrained images.

Attribute	Accuracy	Attribute	Accuracy
Asian	92.32%	Mouth Wide Open	89.63%
Attractive Woman	81.13%	Mustache	91.88%
Baby	90.45%	No Beard	89.53%
Bags Under Eyes	86.23%	No Eyewear	93.55%
Bald	83.22%	Nose Shape	86.87%
Bangs	88.70%	Nose Size	87.50%
Black	88.65%	Nose-Mouth Lines	93.10%
Black Hair	80.32%	Obstructed Forehead	79.11%
Blond Hair	78.05%	Oval Face	70.26%
Blurry	92.12%	Pale Skin	89.44%
Brown Hair	72.42%	Posed Photo	69.72%
Child	83.58%	Receding Hairline	84.15%
Chubby	77.24%	Rosy Cheeks	85.82%
Color Photo	95.50%	Round Face	74.33%
Curly Hair	68.88%	Round Jaw	66.99%
Double Chin	77.68%	Semi-Obscured Forehead	77.02%
Environment	84.80%	Senior	88.74%
Eye Width	90.02%	Shiny Skin	84.73%
Eyebrow Shape	80.90%	Sideburns	71.07%
Eyebrow Thickness	93.40%	Smiling	95.33%
Eyeglasses	91.56%	Soft Lighting	67.81%
Eyes Open	92.52%	Square Face	81.19%
Flash Lighting	72.33%	Straight Hair	76.81%
Frowning	95.47%	Sunglasses	94.91%
Goatee	80.35%	Teeth Not Visible	91.64%
Gray Hair	87.18%	Teeth Visible	91.64%
Harsh Lighting	78.74%	Visible Forehead	89.43%
High Cheekbones	84.70%	Wavy Hair	64.49%
Indian	86.47%	Wearing Hat	85.97%
Male	81.22%	Wearing Lipstick	86.78%
Middle-Aged	78.39%	White	91.48%
Mouth Closed	89.27%	Youth	85.79%
Mouth Partially Open	85.13%

High classification accuracy ( $>90\%$ ) is attained on distinct high-contrast attributes including Color Photo ( $95.50\%$ ), Frowning ( $95.47\%$ ), Smiling ( $95.33\%$ ), Sunglasses ( $94.91\%$ ), and No Eyewear ( $93.55\%$ ), whereas subtle texture traits such as Wavy Hair ( $64.49\%$ ) and Soft Lighting ( $67.81\%$ ) exhibit lower individual accuracy.

Coverage note — None was omitted; all contributed models, dataset definitions, formulations, and empirical findings are fully represented.

References

1.M. S. Bartlett, G. Littlewort, I. Fasel, and J. R. Movellan. Real time face detection and facial expression recognition: Development and applications to human computer interaction. CVPRW, 05, 2003.
2.T. L. Berg, A. C. Berg, J. Edwards, M. Maire, R. White, Y.-W. Teh, E. Learned-Miller, and D. A. Forsyth. Names and faces in the news. CVPR 2004.
3.V. Blanz, S. Romdhani, and T. Vetter. Face identification across different poses and illuminations with a 3d morphable model. FGR, 2002.
4.C. D. Castillo and D. W. Jacobs. Using stereo matching for 2-d face recognition across pose. CVPR, 2007.
5.C.-C. Chang and C.-J. Lin. LIBSVM: a library for support vector machines. http://www.csie.ntu.edu.tw/~cjlin/libsvm, 2001.
6.H. F. Chen, P. N. Belhumeur, and D. W. Jacobs. In search of illumination invariants. CVPR, 2000.
7.T. Cootes, K. Walker, and C. Taylor. View-based active appearance models. FGR 2000.
8.C. Cortes and V. Vapnik. Support-vector networks. Machine Learning, 20(3), 1995.
9.G. W. Cottrell and J. Metcalfe. Empath: face, emotion, and gender recognition using holons. In NIPS, pages 564–571, 1990.
10.M. Everingham, J. Sivic, and A. Zisserman. Hello! my name is... Buffy – automatic naming of characters in TV video. BMVC 2006.
11.A. Ferencz, E. Learned-Miller, and J. Malik. Learning to locate informative features for visual identification. IJCV Spec. Issue on Learning and Vision, 2007.
12.Y. Freund and R. E. Schapire. Experiments with a new boosting algorithm. ICML, 1996.
13.A. Gallagher and T. Chen. Estimating age, gender, and identity using first name priors. CVPR, 2008.
14.A. S. Georghiades, P. N. Belhumeur, and D. J. Kriegman. From few to many: Illumination cone models for face recognition under variable lighting and pose. PAMI, 23(6):643–660, 2001.
15.B. A. Golomb, D. T. Lawrence, and T. J. Sejnowski. SexNet: A neural network identifies sex from human faces. In NIPS, pages 572–577, 1990.
16.R. Gross, J. Shi, and J. Cohn. Quo vadis face recognition? In Workshop on Empirical Evaluation Methods in Computer Vision, December 2001.
17.G. Huang, V. Jain, and E. Learned-Miller. Unsupervised joint alignment of complex images. ICCV, 2007.
18.G. Huang, M. Jones, and E. Learned-Miller. LFW results using a combined Nowak plus MERL recognizer. 2008.
19.G. Huang, M. Ramesh, T. Berg, and E. Learned-Miller. Labeled Faces in the Wild: A database for studying face recognition in unconstrained environments. UMass Amherst Technical Report 07-49, October 2007.
20.J. Huang, X. Shao, and H. Wechsler. Face pose discrimination using support vector machines (SVM). ICPR, pages 154–156, 1998.
21.N. Kumar, P. N. Belhumeur, and S. K. Nayar. FaceTracer: A search engine for large collections of images with faces. ECCV, 2008.
22.H. Ling, S. Soatto, N. Ramanathan, and D. Jacobs. A study of face recognition as people age. ICCV, 2007.
23.D. Lowe. Distinctive image features from scale-invariant keypoints. IJCV, 2003.
24.B. Moghaddam and M.-H. Yang. Learning gender with support faces. PAMI, 24(5):707–711, May 2002.
25.E. Nowak and F. Jurie. Learning visual similarity measures for comparing never seen objects. CVPR, 2007.
26.Omron. OKAO vision. http://www.omron.com/r_d/coretech/vision/okao.html, 2009.
27.A. O'Toole, P. Phillips, F. Jiang, J. Ayyad, N. Penard, and H. Abdi. Face recognition algorithms surpass humans matching faces over changes in illumination. PAMI, 29(9):1642–1646, Sept. 2007.
28.A. Pentland, B. Moghaddam, and T. Starner. View-based and modular eigenspaces for face recognition. CVPR, 1994.
29.P. Phillips, P. Flynn, T. Scruggs, K. Bowyer, and W. Worek. Preliminary face recognition grand challenge results. FGR 2006, pages 15–24.
30.P. Phillips, H. Moon, S. Rizvi, and P. Rauss. The FERET evaluation methodology for face-recognition algorithms. PAMI, 22(10):1090–1104, 2000.
31.F. Samaria and A. Harter. Parameterisation of a stochastic model for human face identification. Workshop on Applications of Computer Vision, 1994.
32.G. Shakhnarovich, P. Viola, and B. Moghaddam. A unified learning framework for real time face detection and classification. FGR, 2002.
33.T. Sim, S. Baker, and M. Bsat. The CMU Pose, Illumination, and Expression (PIE) database. FGR, 2002.
34.L. Wolf, T. Hassner, and Y. Taigman. Descriptor based methods in the wild. In Real-Life Images workshop at ECCV, 2008.
35.W. Zhao, R. Chellappa, P. J. Phillips, and A. Rosenfeld. Face recognition: A literature survey. ACM Comput. Surv., 35(4):399–458, 2003.

Citation

MLA

Kumar, N., et al. “Attribute and Simile Classifiers for Face Verification”. 2009 IEEE 12th International Conference on Computer Vision, 2009, pp. 365–72, https://doi.org/10.1109/ICCV.2009.5459250.

APA

Kumar, N., Berg, A. C., Belhumeur, P. N., & Nayar, S. K. (2009). Attribute and simile classifiers for face verification. 2009 IEEE 12th International Conference on Computer Vision, 365–372. https://doi.org/10.1109/ICCV.2009.5459250

Chicago

Kumar, N., A. C. Berg, P. N. Belhumeur, and S. K. Nayar. 2009. “Attribute and Simile Classifiers for Face Verification”. 2009 IEEE 12th International Conference on Computer Vision, 365–72. https://doi.org/10.1109/ICCV.2009.5459250.

Harvard

Kumar, N. et al. (2009) “Attribute and simile classifiers for face verification”, 2009 IEEE 12th International Conference on Computer Vision. IEEE, pp. 365–372. Available at: https://doi.org/10.1109/ICCV.2009.5459250.

Vancouver

1. Kumar N, Berg AC, Belhumeur PN, Nayar SK (2009) Attribute and simile classifiers for face verification. In: 2009 IEEE 12th International Conference on Computer Vision. IEEE, pp 365–372

BibTeX

@inproceedings{Kumar_2009, title={Attribute and simile classifiers for face verification}, url={http://dx.doi.org/10.1109/ICCV.2009.5459250}, DOI={10.1109/iccv.2009.5459250}, booktitle={2009 IEEE 12th International Conference on Computer Vision}, publisher={IEEE}, author={Kumar, Neeraj and Berg, Alexander C and Belhumeur, Peter N and Nayar, Shree K}, year={2009}, month=Sept, pages={365–372} }

Metadata:Crossref

Access the Paper

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

Open PDF

License: IEEE

Attribute and simile classifiers for face verification

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Trait-Based Face Verification Decision Pipeline

Knowl 2 — Reference-Based Simile Classifiers for Facial Representation

Knowl 3 — Semantic Attribute Classifiers for Facial Representation

Knowl 4 — Face Verification Performance on the Labeled Faces in the Wild Benchmark

Knowl 5 — Human Verification Accuracy and Background Context Bias on LFW

Knowl 6 — Low-Level Facial Feature Extraction and Normalization Pipeline

Knowl 7 — Public Figures (PubFig) Dataset and Benchmark Protocol

Knowl 8 — PubFig Verification Performance across Pose, Illumination, and Expression Variations

Knowl 9 — Accuracies of 65 Semantic Facial Attribute Classifiers

References

Citation

BibTeX

Access the Paper