One-shot learning of object categories

Li Fei-FeiR. FergusP. Perona

article2006TPAMI3,365 citations

Develops a Bayesian framework that leverages prior knowledge from previously learned classes to recognize novel visual object categories from as little as a single training image.

Listen

Learning visual categories of objects has long required hundreds or thousands of labeled training images per category, along with manual alignment of examples, making it impractical to build systems that recognize the tens of thousands of categories humans handle. The paper addresses this bottleneck by testing whether prior knowledge from a small number of already-learned categories can enable useful models of new categories from just one or a few images.

The work set out to evaluate a Bayesian method that represents object categories as probabilistic constellation models of shape and appearance and encodes knowledge from earlier categories as a prior distribution over model parameters. The prior is updated with limited new observations to produce a posterior that supports detection and recognition. Experiments used weakly supervised training on four standard categories and then on a 101-category collection assembled from web images, with performance measured by equal-error rates on held-out test images against background scenes. The Bayesian approach was compared directly with maximum-likelihood and maximum-a-posteriori baselines using identical data splits and repeated random trials.

With one training image the Bayesian method reached roughly 70–85 percent detection on several categories where the other two methods could not produce a model at all; with three to six images it typically reached 80–95 percent on well-represented categories and remained above chance on most of the 101. The same prior, derived from only three unrelated categories, improved results across widely varying objects, and full shape-plus-appearance models outperformed either cue alone. Discrimination among all 101 categories rose from about 10 percent correct with three examples to roughly 18 percent with fifteen—still far above the 1 percent random baseline.

These results show that prior knowledge can substitute for large training sets, lowering the cost and supervision needed to expand visual recognition systems and bringing machine performance closer to the rapid category acquisition observed in children. Because the method works with unaligned web images, it also reduces the practical barrier to scaling recognition to thousands of categories.

The main limitations are the use of a simple prior built from only three source categories, a four-part single-viewpoint model that omits occlusion, and uneven feature detection quality across categories. Performance therefore varies and remains below the level needed for many applications. Further gains are likely from richer priors learned from dozens of categories, incremental updating of the prior as new categories arrive, and modest increases in model complexity.

California Institute of Technology University of Illinois Urbana-Champaign University of Oxford maximum likelihood one-shot learning probabilistic models One‑Shot Learning Bayesian Approach Object Category

Fei-Fei et al (2006).pdf

Paper: A Bayesian hierarchical model for learning natural scene categories, Li Fei-Fei et al. (2005). This earlier hierarchical scene categorization work establishes the foundational Bayesian modeling framework and codebook ideas adapted for one-shot learning across diverse categories.

Paper: Matching Networks for One Shot Learning, Oriol Vinyals et al. (2016). Building directly on the one-shot learning paradigm, this work advances the field by introducing metric spaces and attention mechanisms for rapid classification without parameter updates.
Paper: Prototypical Networks for Few-shot Learning, Jake Snell et al. (2017). This paper extends few-shot classification concepts into simpler, highly effective embedding spaces defined by class prototypes.
Paper: DeViSE: A Deep Visual-Semantic Embedding Model, Andrea Frome et al. (2013). Extending beyond visual-only priors, this research applies embedding models to bridge visual and semantic spaces for zero-shot object recognition.

Cover for One-shot learning of object categories

Abstract

Abstract—Learning visual models of object categories notoriously requires hundreds or thousands of training examples. We show that it is possible to learn much information about a category from just one, or a handful, of images. The key insight is that, rather than learning from scratch, one can take advantage of knowledge coming from previously learned categories, no matter how different these categories might be. We explore a Bayesian implementation of this idea. Object categories are represented by probabilistic models. Prior knowledge is represented as a probability density function on the parameters of these models. The posterior model for an object category is obtained by updating the prior in the light of one or more observations. We test a simple implementation of our algorithm on a database of 101 diverse object categories. We compare category models learned by an implementation of our Bayesian approach to models learned from Maximum Likelihood (ML) and Maximum A Posteriori (MAP) methods. We find that on a database of more than 100 categories, the Bayesian approach produces informative models when the number of training examples is too small for other methods to operate successfully.

1 INTRODUCTION
2 LITERATURE REVIEW
3 THEORETICAL APPROACH
3.1 Overall Bayesian Framework
3.2 The Object Category Model
3.2.1 Appearance
3.2.2 Shape
3.3 Discussion of the Model
3.4 Form of the Parameter Posterior
3.4.1 Maximum Likelihood (ML) and Maximum A Posteriori (MAP)
3.4.2 Other Inference Methods
3.4.3 Conjugate Densities
3.5 Recognition Using a Conjugate Density Parameter Posterior
3.5.1 Parameter Distribution
3.5.2 Closed-Form Calculation of R
4 LEARNING USING A CONJUGATE DENSITY PARAMETER POSTERIOR
5 IMPLEMENTATION
5.1 Feature Detection and Representation
5.2 Learning
5.2.1 Choice of Prior
5.2.2 Details of the Bayesian One-Shot Algorithm
6 EXPERIMENTAL RESULTS
6.1 Data Sets
6.2 Experimental Setup
6.3 Walkthrough for the Motorbike Category
6.4 Caltech 4 Data Set
6.5 Caltech 101 Data Set
6.5.1 Overall Results: ML versus MAP versus Bayesian
6.5.2 Good Models and Bad Models
6.5.3 A Further Investigation on Prior Models and Feature Detectors
6.5.4 Bayesian One-Shot Algorithm: Shape-Only versus App-Only versus Shape-App models
6.5.5 Bayesian One-Shot Algorithm: Discrimination among 101 Categories
6.5.6 Discussions
7 CONCLUSIONS AND FUTURE WORK

Knowls

Knowl 1 — Probabilistic Constellation Model for Object Categories
model/method
The constellation model represents an object category as a constellation of $P$ parts with flexible geometry and distinctive appearance. Given an image $I$ , an interest point detector extracts $N$ candidate regions with spatial locations $X = [x_1, \dots, x_N]$ and appearances $A = [a_1, \dots, a_N] \in \mathbb{R}^{k \times N}$ (where $k$ is the dimensionality of PCA-reduced appearance patches, set to $k = 10$ ).

A hypothesis $h = (h_1, \dots, h_P)$ is a vector assigning each of the $P$ model parts to one of the $N$ detected image features ( $1 \le h_p \le N$ ). Unassigned features are treated as background clutter. An ordering constraint (e.g., requiring the $x$ -coordinates of parts in $h$ to be strictly monotonically increasing) reduces the number of candidate hypotheses $|H|$ by $P!$ .

Assuming shape $X$ and appearance $A$ are conditionally independent given hypothesis $h$ and model parameters $\theta = \{\pi, \theta^A, \theta^X\}$ for a mixture of $\Omega$ components: $p(X, A | \theta) = \sum_{\omega=1}^{\Omega} p(\omega | \theta) \sum_{h \in H} p(A | h, \theta_\omega^A) p(X | h, \theta_\omega^X) p(h | \omega)$
1. Appearance Model: For mixture component $\omega$ , the appearance of part $p$ is modeled as a Gaussian $\mathcal{N}(A(h_p) | \mu_{p,\omega}^A, (\Gamma_{p,\omega}^A)^{-1})$ with diagonal precision matrix $\Gamma_{p,\omega}^A$ . Background features follow a fixed Gaussian $\mathcal{N}(A(j) | \mu_{\text{bg}}^A, (\Gamma_{\text{bg}}^A)^{-1})$ . Evaluating the ratio of assigned foreground to background features yields: $\frac{p(A | h, \theta_\omega^A)}{p(A | h_0, \theta_{\text{bg}}^A)} = \prod_{p=1}^P \frac{\mathcal{G}(A(h_p) | \mu_{p,\omega}^A, \Gamma_{p,\omega}^A)}{\mathcal{G}(A(h_p) | \mu_{\text{bg}}^A, \Gamma_{\text{bg}}^A)}$ where $\mathcal{G}(v | \mu, \Gamma)$ denotes a Gaussian density in $v$ with mean $\mu$ and precision $\Gamma$ , and $h_0$ denotes the null hypothesis assigning all features to background.
2. Shape Model: Features are transformed into a translation- and scale-invariant representation relative to a landmark part (the leftmost part). The joint relative positions of the remaining $P-1$ parts are modeled as a $2(P-1)$ -dimensional Gaussian with mean $\mu_\omega^X$ and full precision matrix $\Gamma_\omega^X$ . Background feature positions are modeled uniformly over image area $\alpha$ with density $\alpha^{-1}$ : $\frac{p(X | h, \theta_\omega^X)}{p(X | h_0, \theta_{\text{bg}}^X)} = \alpha^{P-1} \mathcal{G}(X(h) | \mu_\omega^X, \Gamma_\omega^X)$
Knowl 2 — Conjugate Parameter Prior for the Constellation Model
model/method
To enable analytical integration over parameter uncertainty in few-shot settings, conjugate priors are assigned to the parameters $\theta = \{\pi, \boldsymbol{\mu}^X, \boldsymbol{\mu}^A, \boldsymbol{\Gamma}^X, \boldsymbol{\Gamma}^A\}$ of the constellation model:

$p(\theta) = p(\pi) \prod_{\omega=1}^\Omega p(\boldsymbol{\mu}_\omega^X | \boldsymbol{\Gamma}_\omega^X) p(\boldsymbol{\Gamma}_\omega^X) \prod_{p=1}^P p(\boldsymbol{\mu}_{p,\omega}^A | \boldsymbol{\Gamma}_{p,\omega}^A) p(\boldsymbol{\Gamma}_{p,\omega}^A)$

where:
1. Mixing proportions: Distributed as a symmetric Dirichlet density, $p(\pi) = \text{Dir}(\lambda I)$ .
2. Precision matrices: Distributed as Wishart densities, $p(\boldsymbol{\Gamma}^X) = \mathcal{W}(\boldsymbol{\Gamma}^X | a^X, \mathbf{B}^X)$ and $p(\boldsymbol{\Gamma}^A) = \mathcal{W}(\boldsymbol{\Gamma}^A | a^A, \mathbf{B}^A)$ , where $a$ denotes degrees of freedom and $\mathbf{B}$ is a scale matrix.
3. Mean vectors conditioned on precision: Distributed as Gaussian densities, $p(\boldsymbol{\mu}^X | \boldsymbol{\Gamma}^X) = \mathcal{N}(\boldsymbol{\mu}^X | \mathbf{m}^X, (\beta^X \boldsymbol{\Gamma}^X)^{-1})$ and $p(\boldsymbol{\mu}^A | \boldsymbol{\Gamma}^A) = \mathcal{N}(\boldsymbol{\mu}^A | \mathbf{m}^A, (\beta^A \boldsymbol{\Gamma}^A)^{-1})$ , where $\mathbf{m}$ is the prior mean and $\beta$ is a scalar confidence parameter.
The joint distribution over each mean and precision pair $(\boldsymbol{\mu}, \boldsymbol{\Gamma})$ is a Normal-Wishart conjugate density $\mathcal{NW}(\mathbf{m}, \beta, a, \mathbf{B})$ . For $P=4$ parts and $k=10$ PCA appearance dimensions, the model contains 107 parameters and 109 hyperparameters.
Knowl 3 — Closed-Form Marginal Likelihood via Student's T Distribution
theoretical result
When parameter distributions are conjugate Normal-Wishart densities, the marginal likelihood of observing features $(X, A)$ given training data $(X_t, A_t)$ is obtained by integrating out the model parameters $\theta$ analytically:

$p(X, A | X_t, A_t, O_{\text{fg}}) = \int p(X, A | \theta) p(\theta | X_t, A_t, O_{\text{fg}}) \, d\theta$

Evaluating this integral transforms the Gaussian likelihoods into a mixture of multimodal multivariate Student's T distributions $\mathcal{S}$ :

$p(X, A | X_t, A_t, O_{\text{fg}}) = \sum_{\omega=1}^\Omega \sum_{h=1}^{|H|} \tilde{\pi}_\omega \mathcal{S}\left(X(h) \,\middle|\, \gamma_\omega^X, \mathbf{m}_\omega^X, \boldsymbol{\Lambda}_\omega^X\right) \mathcal{S}\left(A(h) \,\middle|\, \gamma_\omega^A, \mathbf{m}_\omega^A, \boldsymbol{\Lambda}_\omega^A\right)$

where $d$ is the dimensionality of the corresponding parameter vector, and the updated parameters are defined as: $\gamma_\omega = a_\omega + 1 - d$ $\boldsymbol{\Lambda}_\omega = \frac{\beta_\omega + 1}{\beta_\omega \gamma_\omega} \mathbf{B}_\omega$ $\tilde{\pi}_\omega = \frac{\lambda_\omega}{\sum_{\omega'} \lambda_{\omega'}}$

Detection is performed using the posterior odds ratio $R$ : $R = \frac{p(X, A | X_t, A_t, O_{\text{fg}})}{p(X, A | \theta_{\text{bg}})}$ where the background parameter $\theta_{\text{bg}}$ is estimated once via Maximum Likelihood from a generic background dataset, collapsing the denominator to a deterministic baseline.
Knowl 4 — Estimation of Object Priors from Previously Learned Categories
algorithm
Prior hyperparameters $(\mathbf{m}_0, \beta_0, a_0, \mathbf{B}_0)$ are computed empirically from $M$ Maximum Likelihood (ML) models trained on previously learned distinct object categories (e.g., $M = 30$ models obtained from spotted cats, faces, and airplanes):
Input: Set of M maximum-likelihood models with parameters {mu_m^ML, Gamma_m^ML} for m = 1 to M
Input: Number of parts P, appearance dimension k
Output: Prior hyperparameters (m_0^X, beta_0^X, a_0^X, B_0^X) and (m_0^A, beta_0^A, a_0^A, B_0^A)
1. Set degrees of freedom a_0:
   a_0^X = 2 * (P - 1) * (P - 2)
   a_0^A = k * P
2. Compute prior mean vector m_0:
   m_0 = (1 / M) * sum_{m=1}^M mu_m^ML
3. Compute scale matrix B_0:
   Mean_Gamma = (1 / M) * sum_{m=1}^M Gamma_m^ML
   B_0 = a_0 * (Mean_Gamma)^{-1}
4. Compute precision scaling factor beta_0 as the ratio of the precision of the means to the mean of precisions:
   Var_mu = (1 / M) * sum_{m=1}^M (mu_m^ML - m_0)^2
   beta_0 = || Var_mu ||^{-1} / || a_0 * (B_0)^{-1} ||
return (m_0, beta_0, a_0, B_0) for both shape (X) and appearance (A)
This captures general visual properties (e.g., that parts tend to form coherent spatial clusters and localized appearance variances) without requiring visual similarity between the prior categories and new target classes.
Knowl 5 — Caltech-101 Dataset and Evaluation Setup
experimental setup
The Caltech-101 dataset consists of 101 diverse foreground object categories plus one background clutter category (obtained from Google Image Search using the query word 'things'). Category names were chosen from Webster's Collegiate Dictionary entries that had illustrations. Each category contains between 41 and 800 images (typically 45 to 400), evaluated in grayscale.

Key experimental protocol details:
- Preprocessing: Images with horizontal mirror ambiguity were flipped to face one direction; vertically structured categories were rotated.
- Feature Extraction: Salient region detector (Kadir & Brady) extracts scale-invariant regions (average 20 interest points per image). Regions are rescaled to $11 \times 11$ patches and projected onto the top 10 principal components precalculated from background data.
- Model Parameters: $P = 4$ parts, single mixture component $\Omega = 1$ .
- Evaluation Protocol: Each category is split into equal train and test sets. $N$ training images ( $N \in \{1, 2, 3, 4, 5, 6, 15\}$ ) are sampled randomly across 10 independent trials. The test set consists of 50 foreground images and 50 background images. Performance is measured using the Equal Error Rate (EER) on the Receiver Operating Characteristic (ROC) curve (where $p(\text{True Positive}) = 1 - p(\text{False Alarm})$ ).
Knowl 6 — Few-Shot Categorization Performance: Bayesian vs. ML vs. MAP
empirical result
Experimental comparison of Bayesian One-Shot learning against Maximum Likelihood (ML) and Maximum A Posteriori (MAP) across the Caltech-101 dataset demonstrates:
1. Singular Example Regimes ( $N=1$ ): ML and MAP fail to converge due to parameter singularity. The Bayesian One-Shot method achieves an average detection performance of ~70% to 95% equal error rate across categories with a single training image (e.g., 82% on faces with $N=1$ ).
2. Underfitting vs. Overfitting: MAP incorporates prior knowledge but computes a point estimate $\theta^{\text{MAP}} = \arg\max_\theta p(X_t, A_t | \theta) p(\theta)$ , which overfits when $N \le 6$ . By integrating over the entire parameter posterior density $p(\theta | X_t, A_t)$ , the Bayesian algorithm accounts for parameter uncertainty and consistently outperforms both ML and MAP.
3. Sample Efficiency: On categories such as grand piano and cougar face, Bayesian One-Shot reaches equal error rates of 84% and 85% at $N=15$ , whereas ML error rates are substantially higher when training numbers are small ( $N \le 6$ ).
Knowl 7 — Multiclass Discrimination across 101 Object Categories
empirical result
In a 101-way multiclass discrimination task, query images from each category's test set are evaluated against all 101 learned category models, assigning class labels via winner-take-all maximum likelihood score over 50 random test iterations.

The mean diagonal classification accuracy across all 101 categories:
- $N = 3$ training examples: $10.4\%$
- $N = 6$ training examples: $13.9\%$
- $N = 15$ training examples: $17.7\%$
For comparison, random chance on the 101-class problem is $\approx 1.0\%$ (1/101), indicating significant discriminatory category information learned from very few exemplars without category-specific manual tuning.
Knowl 8 — Ablation of Shape and Appearance Components in Object Models
empirical result
Evaluating lesioned models (shape-only and appearance-only) against the full joint shape-appearance constellation model across all 101 categories reveals differing class dependencies:
1. Shape-Dominant Categories: Objects with rigid geometric configurations but varied surface textures (e.g., human faces, electric guitars, side-view cars) rely predominantly on the shape density term.
2. Appearance-Dominant Categories: Objects with highly characteristic local textures but deformable or unconstrained geometry (e.g., leopards, octopi, ketch) achieve higher performance with appearance-only models.
3. Combined Superiority: Across all tested sample sizes ( $N \in \{1, 3, 6, 15\}$ ), the combined shape-appearance model systematically outperforms both shape-only and appearance-only models.
Knowl 9 — Effect of Prior Category Diversity and Prior Specificity
empirical result
Empirical analysis of prior construction dynamics demonstrates:
1. Number of Prior Categories: Incrementally increasing the number of pre-learned categories used to construct the prior from $N_{\text{prior}} = 1$ to 20 over a test set of 30 unseen categories shows that recognition error steadily decreases as $N_{\text{prior}}$ increases from 1 to 8, with performance gains plateauing beyond approximately 8 categories.
2. Prior Relevance and ML Convergence: Synthetic experiments using 4-part geometric models (triangles, trapeziums, squares) show that when the prior matches the true underlying class structure (e.g., triangular prior for learning triangular shapes), the Bayesian model maintains a significant performance advantage over Maximum Likelihood for more than 100 training examples. With generic or weakly matching priors, ML catches up within 30 training examples.
Knowl 10 — Structural and Feature Bottlenecks of the Bayesian Constellation Formulation
limitation
The simplified Bayesian constellation model has several operational limitations:
1. Computational Bottleneck of Part Matching: Modeling joint shape covariance over all parts requires evaluating $O(N^P)$ hypotheses per image. Without heuristic search or tree-structured conditional independence, the model is practically restricted to $P \le 4$ parts.
2. Dependence on Salient Feature Detectors: The model generates features rather than raw pixels. If the interest point detector fails to fire on the foreground object (e.g., categories like crocodile and mayfly with weak boundary saliency), recognition error increases significantly; correlation exists between the percentage of detections falling within object boundaries and classification performance.
3. Viewpoint and Articulation Rigidity: Using a single mixture component ( $\Omega = 1$ ) restricts the model to a single dominant viewpoint and compact, non-articulated objects.
4. Omission of Occlusion Statistics: To ensure closed-form variational tractability with few images, explicit occlusion modeling and feature detector missed-detection statistics present in earlier constellation formulations were omitted.

Coverage note — Omitted mathematical derivations of the Variational Bayesian Expectation-Maximization (VBEM) algorithm and MAP updates as they are standard variational calculus derivations deferred to cited companion technical reports.

References

1.Merriam-Webster’s Collegiate Dictionary, 10th ed., Springfield, Mass.: Merriam-Webster, Inc., 1994.
2.Y. Amit and D. Geman, “A Computational Model for Visual Selection,” Neural Computation, vol. 11, no. 7, pp. 1691-1715, 1999.
3.H. Attias, “Inferring Parameters and Structure of Latent Variable Models by Variational Bayes,” Proc. 15th Conf. Uncertainty in Artificial Intelligence, pp. 21-30, 1999.
4.I. Biederman, “Recognition-by-Components: A Theory of Human Image Understanding,” Psychological Rev., vol. 94, pp. 115-147, 1987.
5.M. Burl and P. Perona, “Recognition of Planar Object Classes,” Proc. Conf. Computer Vision and Pattern Recognition, pp. 223-230, 1996.
6.M. Burl, M. Weber, and P. Perona, “A Probabilistic Approach to Object Recognition Using Local Photometry and Global Geometry,” Proc. European Conf. Computer Vision, pp. 628-641, 1996.
7.A. Dempster, N. Laird, and D. Rubin, “Maximum Likelihood from Incomplete Data via the EM Algorithm,” J. Royal Statistical Soc., vol. 29, pp. 1-38, 1976.
8.L. Fei-Fei, R. Fergus, and P. Perona, “A Bayesian Approach to Unsupervised One-Shot Learning of Object Categories,” Proc. Ninth Int’l Conf. Computer Vision, pp. 1134-1141, Oct. 2003.
9.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,” Proc. Workshop Generative-Model Based Vision, 2004.
10.L. Fei-Fei, R. Fergus, and P. Perona, supplemental material, http://computer.org/tpami/archives.htm, 2006.
11.P. Felzenszwalb and D. Huttenlocher, “Pictorial Structures for Object Recognition,” Int’l J. Computer Vision, vol. 1, pp. 55-79, 2005.
12.P. Felzenszwalb and D. Huttenlocher, “Representation and Detection of Deformable Shapes,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 27, no. 2, pp. 208-220, Feb. 2005.
13.R. Fergus, P. Perona, and A. Zisserman, “Object Class Recognition by Unsupervised Scale-Invariant Learning,” Proc. Computer Vision and Pattern Recognition, pp. 264-271, 2003.
14.R. Fergus, P. Perona, and A. Zisserman, “A Visual Category Filter for Google Images,” Proc. Eighth European Conf. Computer Vision, 2004.
15.R. Fergus, P. Perona, and A. Zisserman, “A Sparse Object Category Model for Efficient Learning and Exhaustive Recognition,” Proc. Computer Vision and Pattern Recognition, 2005.
16.D. Forsyth and A. Zisserman, “Shape from Shading in the Light of Mutual Illumination,” Image and Vision Computing, pp. 42-29, 1990.
17.A. Gelman, J.B. Carlin, H.S. Stern, and D.B. Rubin, Bayesian Data Analysis. Chapman Hall/CRC, 1995.
18.R. Gilks, S. Richardson, and D. Spiegelhalter, Markov Chain Monte Carlo in Practice. Chapman Hall, 1992.
19.R. Gilks and P. Wild, “Adaptive Rejection Sampling for Gibbs Sampling,” Applied Statistics, vol. 41, pp. 337-348, 1992.
20.W. Grimson and D. Huttenlocher, “On the Sensitivity of the Hough Transform for Object Recognition,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 12, no. 3, pp. 255-274, Mar. 1990.
21.K. Humphreys and M. Titterington, “Some Examples of Recursive Variational Approximations for Bayesian Inference,” Advanced Mean Field Methods. M. Opper and D. Saad, eds., MIT Press, 2001.
22.D. Huttenlocher, G. Klanderman, and W. Rucklidge, “Comparing Images Using the Hausdorff Distance,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 15, no. 9, pp. 850-863, Sept. 1993.
23.T. Kadir and M. Brady, “Scale, Saliency and Image Description,” Int’l J. Computer Vision, vol. 45, no. 2, pp. 83-105, 2001.
24.Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-Based Learning Applied to Document Recognition,” Proc. IEEE, vol. 86, no. 11, pp. 2278-2324, Nov. 1998.
25.Y. LeCun, F. Huang, and L. Bottou, “Learning Methods for Generic Object Recognition with Invariance to Pose and Lighting,” Proc. Conf. Computer Vision and Pattern Recognition, 2004.
26.T. Leung, M. Burl, and P. Perona, “Finding Faces in Cluttered Scenes Using Labeled Random Graph Matching,” Proc. Int’l Conf. Computer Vision, pp. 637-644, 1995.
27.D. Lowe, “Object Recognition from Local Scale-Invariant Features,” Proc. Int’l Conf. Computer Vision, pp. 1150-1157, 1999.
28.K. Mikolajczyk and C. Schmid, “An Affine Invariant Interest Point Detector,” Proc. European Conf. Computer Vision, vol. 1, pp. 128-142, 2002.
29.R. Neal and G. Hinton, “A View of the EM Algorithm that Justifies Incremental, Sparse and Other Variants,” Learning in Graphical Models, M.I. Jordan, ed., pp. 355-368, Norwell, Mass.: Kluwer Academic Press, 1998.
30.W. Penny, “Variational Bayes for d-Dimensional Gaussian Mixture Models,” technical report, Univ. College London, 2001.
31.F. Rothganger, S. Lazebnik, C. Schmid, and J. Ponce, “3D Object Modeling and Recognition Using Affine-Invariant Patches and Multi-View Spatial Constraints,” Proc. Computer Vision and Pattern Recognition, pp. 272-280, 2003.
32.H. Rowley, S. Baluja, and T. Kanade, “Neural Network-Based Face Detection,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 20, no. 1, pp. 23-38, Jan. 1998.
33.E. Sali and S. Ullman, “Combining Class-Specific Fragments for Object Classification,” Proc. British Machine Vision Conf., vol. 1, pp. 203-213, 1999.
34.H. Schneiderman and T. Kanade, “A Statistical Approach to 3D Object Detection Applied to Faces and Cars,” Proc. Computer Vision and Pattern Recognition, pp. 746-751, 2000.
35.K. Sung and T. Poggio, “Example-Based Learning for View-Based Human Face Detection,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 20, no. 1, pp. 39-51, Jan. 1998.
36.P. Viola and M. Jones, “Rapid Object Detection Using a Boosted Cascade of Simple Features,” Proc. Computer Vision and Pattern Recognition, vol. 1, pp. 511-518, 2001.
37.P. Viola, M. Jones, and D. Snow, “Detecting Pedestrians Using Patterns of Motion and Appearance,” Proc. Int’l Conf. Computer Vision, pp. 734-741, 2003.
38.M. Weber, W. Einhaeuser, M. Welling, and P. Perona, “Viewpoint-Invariant Learning and Detection of Human Heads,” Proc. Fourth Int’l Conf. Automated Face and Gesture Recognition, pp. 20-27, 2000.
39.M. Weber, M. Welling, and P. Perona, “Unsupervised Learning of Models for Recognition,” Proc. European Conf. Computer Vision, vol. 2, pp. 101-108, 2000.
40.A. Torralba, K.P. Murphy, and W.T. Freeman, Sharing Features: Efficient Boosting Procedures for Multiclass Object Detection,” Proc. IEEE Conf. Computer Vision and Pattern Recognition, pp. 762-769, 2004.
41.M. Weber, “Unsupervised Learning of Models for Object Recognition,” PhD thesis, Calif. Inst. of Technology, Pasadena, 2000.
42.R. Fergus, “Visual Object Category Recognition,” PhD thesis, Univ. of Oxford, U.K., 2005.
43.A. Berg, T. Berg, and J. Malik, “Shape Matching and Object Recognition Using Low Distortion Correspondence,” Proc. IEEE Conf. Computer Vision and Pattern Recognition, pp. 26-33, June 2005.

Citation

MLA

Li Fei-Fei, et al. “One-shot Learning of Object Categories”. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 28, no. 4, 2006, pp. 594–611, https://doi.org/10.1109/TPAMI.2006.79.

APA

Li Fei-Fei, Fergus, R., & Perona, P. (2006). One-shot learning of object categories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(4), 594–611. https://doi.org/10.1109/TPAMI.2006.79

Chicago

Li Fei-Fei, R. Fergus, and P. Perona. 2006. “One-shot Learning of Object Categories”. IEEE Transactions on Pattern Analysis and Machine Intelligence 28 (4): 594–611. https://doi.org/10.1109/TPAMI.2006.79.

Harvard

Li Fei-Fei, Fergus, R. and Perona, P. (2006) “One-shot learning of object categories”, IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(4), pp. 594–611. Available at: https://doi.org/10.1109/TPAMI.2006.79.

Vancouver

1. Li Fei-Fei, Fergus R, Perona P (2006) One-shot learning of object categories. IEEE Transactions on Pattern Analysis and Machine Intelligence 28:594–611

BibTeX

@article{Li_Fei_Fei_2006, title={One-shot learning of object categories}, volume={28}, ISSN={0162-8828}, url={http://dx.doi.org/10.1109/TPAMI.2006.79}, DOI={10.1109/tpami.2006.79}, number={4}, journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, publisher={Institute of Electrical and Electronics Engineers (IEEE)}, author={Li Fei-Fei and Fergus, R. and Perona, P.}, year={2006}, month=Apr, pages={594–611} }

Metadata:Crossref

Access the Paper

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

Open PDF

One-shot learning of object categories

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Probabilistic Constellation Model for Object Categories

Knowl 2 — Conjugate Parameter Prior for the Constellation Model

Knowl 3 — Closed-Form Marginal Likelihood via Student's T Distribution

Knowl 4 — Estimation of Object Priors from Previously Learned Categories

Knowl 5 — Caltech-101 Dataset and Evaluation Setup

Knowl 6 — Few-Shot Categorization Performance: Bayesian vs. ML vs. MAP

Knowl 7 — Multiclass Discrimination across 101 Object Categories

Knowl 8 — Ablation of Shape and Appearance Components in Object Models

Knowl 9 — Effect of Prior Category Diversity and Prior Specificity

Knowl 10 — Structural and Feature Bottlenecks of the Bayesian Constellation Formulation

References

Citation

BibTeX

Access the Paper