Learning To Count Objects in Images

article2010NeurIPS1,372 citations

Develops a supervised object counting framework that bypasses individual instance detection by learning an image density function using dot annotations, optimized with a maximum subarray-based loss for accurate and fast count estimation across image subregions.

Listen

Visual object counting in digital imagery is critical for practical tasks such as biomedical cell analysis, crowd surveillance, wildlife tracking, and forestry surveys. Existing computer vision solutions face severe practical trade-offs. Standard detection methods struggle when objects overlap and require expensive, detailed bounding-box or segmentation annotations. Conversely, regression methods that predict total counts from global image statistics require large volumes of fully labeled images and discard spatial information.

The article demonstrates a supervised learning framework that counts objects by estimating continuous image density maps trained on dot annotations (a single dot per object instance). It evaluates this framework to demonstrate that inferring a regional density function delivers superior counting accuracy compared to existing detection, regression, and application-specific baselines.

The approach converts dot annotations into ground-truth density functions and trains a linear model over image features using a regularized risk framework. To evaluate errors effectively, the authors introduced the Maximum Excess over SubArrays (MESA) distance metric. This metric measures the largest discrepancy between estimated and ground-truth density across all possible rectangular subregions in an image. The optimization problem is solved using a convex quadratic program with cutting-plane methods. The framework was evaluated on two key benchmarks: synthetic fluorescence microscopy datasets of bacterial cell populations (with averages around 171 cells per image) and a real-world 2,000-frame pedestrian surveillance video.

The evaluations yielded several major findings. First, on cell counting, the proposed method substantially outperformed all baselines across all training set sizes; with only one training image, it achieved a mean absolute error of 9.5 cells, compared to 20.8 for detection, 60.4 for kernel regression, and 16.2 for a specialized morphology tool. When trained on 32 images, its error decreased to 3.5 cells. Second, on surveillance footage, the density-based approach delivered mean absolute errors between 1.28 and 2.06 pedestrians across standard test splits, outperforming pure regression methods and matching or exceeding a complex hybrid method that required more detailed supervision. Third, the approach proved robust to local noise and kernel width choices during training, and it introduces almost zero computational overhead during inference beyond feature extraction.

These findings indicate that dot-level spatial supervision offers significant operational benefits. Organizations can drastically lower data labeling costs, as placing a dot per object is faster than drawing bounding boxes or segmenting boundaries. Furthermore, because inference simply involves a linear weighting of extracted image features or random forest paths, the model can run in real time on large datasets, reducing compute infrastructure costs for video surveillance and high-throughput microscopy.

Organizations implementing automated counting pipelines should adopt dot-annotated density estimation where individual object detection is unnecessary or error-prone. For deployment, engineering teams should pair the model with fast feature extractors, such as randomized decision forests, to achieve real-time throughput. To reduce initial training time bottlenecks associated with standard optimization solvers, developers should consider implementing purpose-built solvers tailored to the cutting-plane constraints.

Confidence in these performance results is high for dense stationary and structured video counting problems. However, users should note that the cell experiments relied on synthetic data due to inconsistencies among human annotators on real biological images. In surveillance settings, camera calibration (ground plane depth) was utilized to normalize scale perspective. Deploying this system in novel settings may require validating feature choices and calibration steps under operational conditions.

University of Oxford computer vision feature extraction ground planes object counting density learning cell counting

Lempitsky et al (2010).pdf

Paper: Object Detection with Discriminatively Trained Part-Based Models, Pedro F. Felzenszwalb et al. (2010). Reading this foundational work on discriminatively trained part-based models provides essential context for the traditional sliding-window detection and bounding-box baselines that density-based counting directly addresses and improves upon.
Paper: Region Covariance: A Fast Descriptor for Detection and Classification, Oncel Tuzel et al. (2006). This paper establishes integral image methods for rapid spatial descriptor computation over arbitrary subwindows, which directly inspires fast regional feature extraction techniques used in sliding-window and cutting-plane density estimation.
Paper: TextonBoost: Joint Appearance, Shape and Context Modeling for Multi-class Object Recognition and Segmentation, J. Shotton et al. (2006). This work demonstrates joint appearance, shape, and context modeling using randomized decision trees and spatial filters, providing the foundational feature extraction and random forest principles utilized by the source paper.
Paper: A Trainable System for Object Detection, CONSTANTINE PAPAGEORGIOU et al. (2000). This foundational paper presents the classic sliding-window and regularized margin learning paradigm for object detection that the source contrasts with continuous density map regression.

Paper: Single-Image Crowd Counting via Multi-Column Convolutional Neural Network, Yingying Zhang et al. (2016). This work extends dot-annotated density map estimation to deep learning by introducing multi-column convolutional neural networks and geometry-adaptive kernels for single-image crowd counting.
Paper: CSRNet: Dilated Convolutional Neural Networks for Understanding the Highly Congested Scenes, Yuhong Li et al. (2018). This paper advances spatial crowd density map estimation from dot annotations by employing dilated convolutional networks to generate high-resolution count distributions in highly congested scenes.
Paper: Cell Detection with Star-convex Polygons, Uwe Schmidt et al. (2018). This work continues the problem of dense microscopy cell counting and localization under heavy overlap by predicting star-convex polygons from center probability maps.
Paper: Hover-Net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images, Simon Graham et al. (2018). This paper extends nuclear instance analysis in dense microscopy images by combining central distance regression with simultaneous segmentation and phenotype classification.
Paper: Objects as Points, Xingyi Zhou et al. (2019). This article generalizes point-based spatial density and keypoint regression into a unified deep learning framework that detects objects directly as center points.
Paper: Tracking Objects as Points, Xingyi Zhou et al. (2020). This paper builds on point-centered spatial object localization by tracking multi-object instances as points across video streams.

Cover for Learning To Count Objects in Images

Abstract

We propose a new supervised learning framework for visual object counting tasks, such as estimating the number of cells in a microscopic image or the number of humans in surveillance video frames. We focus on the practically-attractive case when the training images are annotated with dots (one dot per object).

Our goal is to accurately estimate the count. However, we evade the hard task of learning to detect and localize individual object instances. Instead, we cast the problem as that of estimating an image density whose integral over any image region gives the count of objects within that region. Learning to infer such density can be formulated as a minimization of a regularized risk quadratic cost function. We introduce a new loss function, which is well-suited for such learning, and at the same time can be computed efficiently via a maximum subarray algorithm. The learning can then be posed as a convex quadratic program solvable with cutting-plane optimization.

The proposed framework is very flexible as it can accept any domain-specific visual features. Once trained, our system provides accurate object counts and requires a very small time overhead over the feature extraction step, making it a good candidate for applications involving real-time processing or dealing with huge amount of visual data.

Topics

computer vision feature extraction ground planes image features inference linear models mean absolute error object detection optimization optimization problems

Keywords

object counting density learning cell counting object detection cell segmentation SIFT descriptors random forest microscopy images

1 Introduction
1.1 Related work
2 The Framework
2.1 Learning to Count
2.2 The MESA distance
2.3 Optimization
3 Experiments
4 Conclusion
References

Knowls

Knowl 1 — Object Counting via Linear Estimation of Spatial Density Functions
model/method
The counting task is formulated as estimating a continuous/discrete real-valued density function $F: I \to \mathbb{R}$ defined over the pixel grid of an image $I$ . For any image subregion $S \subseteq I$ , the estimated count of objects located within $S$ is obtained by summing the density function over the pixels in that region:

$\hat{C}(S) = \sum_{p \in S} F(p)$

Each pixel $p \in I$ is represented by a feature vector $x_p \in \mathbb{R}^K$ . The density function at pixel $p$ is modeled as a linear transformation of its feature representation:

$F(p \mid w) = w^T x_p$

where $w \in \mathbb{R}^K$ is a learnable parameter vector. Estimating object counts in an unseen test image requires computing the pixel features $x_p$ , taking the inner product with $w$ , and summing the resulting scalar densities over the region of interest. This formulation avoids the discrete classification and non-maximum suppression steps of object detection, while preserving spatial distribution information.
Knowl 2 — Maximum Excess over SubArrays Distance
definition
Given an image pixel grid $I$ and two real-valued density functions $F_1, F_2: I \to \mathbb{R}$ , the Maximum Excess over SubArrays (MESA) distance, denoted $\mathcal{D}_{\mathrm{MESA}}(F_1, F_2)$ , is defined as the maximum absolute difference between the sums of $F_1$ and $F_2$ over all axis-aligned rectangular box subarrays in $I$ :

$\mathcal{D}_{\mathrm{MESA}}(F_1, F_2) = \max_{B \in \mathcal{B}} \left| \sum_{p \in B} F_1(p) - \sum_{p \in B} F_2(p) \right|$

where $\mathcal{B}$ denotes the set of all axis-aligned rectangular box subarrays contained in $I$ .

Equivalently, the distance can be expressed as:

$\mathcal{D}_{\mathrm{MESA}}(F_1, F_2) = \max \left( \max_{B \in \mathcal{B}} \sum_{p \in B} \big(F_1(p) - F_2(p)\big), \; \max_{B \in \mathcal{B}} \sum_{p \in B} \big(F_2(p) - F_1(p)\big) \right)$

The MESA distance operates as an $L_\infty$ metric across subarray integrals. It satisfies two key properties for visual counting:
1. It provides an upper bound on the absolute difference between total counts over the full image (since the whole image $I \in \mathcal{B}$ ).
2. It is robust to zero-mean local perturbations and high-frequency noise (which cancel out over large subarrays) while remaining sensitive to low-frequency shifts in spatial object distributions.
Knowl 3 — Convex Quadratic Program for Density Estimation under MESA Distance
model/method
Given $N$ training images $I_1, \dots, I_N$ , with pixel feature representations $x_p^i \in \mathbb{R}^K$ and target ground truth densities $F^0_i: I_i \to \mathbb{R}$ , the parameter vector $w \in \mathbb{R}^K$ is learned by minimizing an $L_2$ -regularized MESA empirical risk:

$\min_w w^T w + \lambda \sum_{i=1}^N \mathcal{D}_{\mathrm{MESA}}\big(F^0_i(\cdot), F_i(\cdot \mid w)\big)$

where $\lambda > 0$ is a scalar regularization hyperparameter. By introducing an auxiliary non-negative slack variable $\xi_i$ for each training image $i \in \{1, \dots, N\}$ , this optimization is formulated as a convex quadratic program (QP):

$\min_{w, \xi_1, \dots, \xi_N} w^T w + \lambda \sum_{i=1}^N \xi_i$

subject to the linear constraints:

$\forall i \in \{1, \dots, N\}, \; \forall B \in \mathcal{B}_i: \quad \xi_i \ge \sum_{p \in B} \big(F^0_i(p) - w^T x_p^i\big) \quad \text{and} \quad \xi_i \ge \sum_{p \in B} \big(w^T x_p^i - F^0_i(p)\big)$

where $\mathcal{B}_i$ is the set of all axis-aligned box subarrays of image $I_i$ . At the optimal solution $(\hat{w}, \hat{\xi}_1, \dots, \hat{\xi}_N)$ , the slack variables satisfy $\hat{\xi}_i = \mathcal{D}_{\mathrm{MESA}}(F^0_i(\cdot), F_i(\cdot \mid \hat{w}))$ , and $\hat{w}$ solves the regularized risk problem.
Knowl 4 — Cutting-Plane Optimization for MESA-Constrained Density Learning
algorithm
Because the number of box subarray constraints in the convex quadratic program is combinatorial, the problem is solved using an iterative cutting-plane optimization algorithm. At each iteration, the QP is solved over an active constraint subset, and the most violated constraints are identified by solving 2D maximum subarray problems on the residual density maps in $O(|I|^{1.5})$ time using dynamic programming.
Input: Training images $I_1, \dots, I_N$ , pixel features $x_p^i$ , ground truth densities $F^0_i$ , regularization parameter $\lambda > 0$ , tolerance $\epsilon > 0$
Output: Optimal weight vector $w \in \mathbb{R}^K$
Initialize active constraint set $\mathcal{C}$ with a small random subset of box subarrays
repeat
    Solve QP over active constraints $\mathcal{C}$ :
         $(w, \xi_1, \dots, \xi_N) \leftarrow \arg\min_{w, \xi} \left( w^T w + \lambda \sum_{i=1}^N \xi_i \right)$ subject to constraints in $\mathcal{C}$
    violations_found $\leftarrow$ false
    for $i = 1$ to $N$ do
        Compute residual maps $R_1(p) = F^0_i(p) - w^T x_p^i$ and $R_2(p) = w^T x_p^i - F^0_i(p)$
        Find box $B_1^* = \arg\max_{B \in \mathcal{B}_i} \sum_{p \in B} R_1(p)$ via 2D max subarray
        Find box $B_2^* = \arg\max_{B \in \mathcal{B}_i} \sum_{p \in B} R_2(p)$ via 2D max subarray
        if $\sum_{p \in B_1^*} R_1(p) > \xi_i \cdot (1 + \epsilon)$ then
            Add constraint $\xi_i \ge \sum_{p \in B_1^*} (F^0_i(p) - w^T x_p^i)$ to $\mathcal{C}$
            violations_found $\leftarrow$ true
        end if
        if $\sum_{p \in B_2^*} R_2(p) > \xi_i \cdot (1 + \epsilon)$ then
            Add constraint $\xi_i \ge \sum_{p \in B_2^*} (w^T x_p^i - F^0_i(p))$ to $\mathcal{C}$
            violations_found $\leftarrow$ true
        end if
    end for
until violations_found is false
return $w$
Knowl 5 — Kernel Density Ground Truth Generation from Dot Annotations
equation
For a training image $I_i$ annotated with a set of $C(i)$ 2D point coordinates $\mathcal{P}_i = \{P_1, P_2, \dots, P_{C(i)}\}$ , the ground truth continuous density function $F^0_i(p)$ is defined as a sum of normalized 2D Gaussian kernels centered at each annotated dot:

$\forall p \in I_i, \quad F^0_i(p) = \sum_{P \in \mathcal{P}_i} \mathcal{N}(p; P, \sigma^2 I_{2\times 2})$

where $\mathcal{N}(p; P, \sigma^2 I_{2\times 2})$ is a normalized 2D isotropic Gaussian kernel evaluated at pixel coordinate $p$ , with mean $P$ and covariance matrix $\sigma^2 I_{2\times 2}$ . The standard deviation $\sigma$ is a small spatial bandwidth parameter (or $\sigma = 0$ , reducing $F^0_i$ to a sum of discrete delta functions).

When an annotated object is near the image boundary, part of the Gaussian probability mass falls outside the pixel grid $I_i$ , causing $\sum_{p \in I_i} F^0_i(p) < C(i)$ . This boundary truncation directly models partially visible boundary objects as fractional object counts.
Knowl 6 — Multi-Modal Random Forest Embedding and Perspective Scaling for Video Counting
model/method
For counting objects in video sequences with perspective variation (e.g. pedestrians), pixel feature vectors are extracted by combining multiple modalities through a randomized decision forest:
1. Primary pixel features are computed at each pixel $p$ : raw intensity, absolute temporal difference from the previous frame, background subtraction difference (using median filtering for static background estimation), and absolute spatial derivatives in $x$ and $y$ .
2. An ensemble of randomized regression trees (e.g. 5 trees) is trained to map local multi-channel patch appearances to local ground truth densities.
3. Each pixel $p$ is mapped to a binary indicator feature vector $x_p \in \{0, 1\}^K$ , where $K$ is the total number of leaf nodes across all trees in the forest. The vector contains 1 at the indices corresponding to the leaf reached in each tree, and 0 elsewhere.
4. To compensate for perspective scaling, the feature vector $x_p$ is multiplied by the squared ground plane depth $d(p)^2$ at pixel location $p$ :
$x_p \leftarrow d(p)^2 \cdot x_p$

At test time, the density estimate $F(p \mid w) = w^T x_p$ is computed in real time by routing pixel $p$ down each tree, retrieving the scalar weight $w_t$ assigned to each reached leaf node $t$ , and summing the weights scaled by $d(p)^2$ .

Knowl 7 — Bacterial Cell Counting Accuracy across Training Set Sizes

data/table

Mean absolute errors (MAE) on a test set of 100 synthetic fluorescence microscopy images (cell counts ranging from 74 to 317 per image, mean $171 \pm 64$ ) as a function of the number of training images $N \in \{1, 2, 4, 8, 16, 32\}$ (with an equal number of validation images $N$ , averaged over 5 random splits):

Method	Validation Metric	$N=1$	$N=2$	$N=4$	$N=8$	$N=16$	$N=32$
Linear ridge regression	counting	$67.3 \pm 25.2$	$37.7 \pm 14.0$	$16.7 \pm 3.1$	$8.8 \pm 1.5$	$6.4 \pm 0.7$	$5.9 \pm 0.5$
Kernel ridge regression	counting	$60.4 \pm 16.5$	$38.7 \pm 17.0$	$18.6 \pm 5.0$	$10.4 \pm 2.5$	$6.0 \pm 0.8$	$5.2 \pm 0.3$
Detection (SVM + NMS)	counting	$28.0 \pm 20.6$	$20.8 \pm 5.8$	$13.6 \pm 1.5$	$10.2 \pm 1.9$	$10.4 \pm 1.2$	$8.5 \pm 0.5$
Detection (SVM + NMS)	detection	$20.8 \pm 3.8$	$20.1 \pm 5.5$	$15.7 \pm 2.0$	$15.0 \pm 4.1$	$11.8 \pm 3.1$	$12.0 \pm 0.8$
Detection + linear correction	counting	–	$22.6 \pm 5.3$	$16.8 \pm 6.5$	$6.8 \pm 1.2$	$6.1 \pm 1.6$	$4.9 \pm 0.5$
Density learning (proposed)	counting	$12.7 \pm 7.3$	$7.8 \pm 3.7$	$5.0 \pm 0.5$	$4.6 \pm 0.6$	$4.2 \pm 0.4$	$3.6 \pm 0.2$
Density learning (proposed)	MESA	$9.5 \pm 6.1$	$6.3 \pm 1.2$	$4.9 \pm 0.6$	$4.9 \pm 0.7$	$3.8 \pm 0.2$	$3.5 \pm 0.2$

An application-specific baseline using adaptive thresholding and morphological analysis achieved an MAE of $16.2$ . The proposed MESA-based density learning outperforms global regression baselines and object detection baselines across all sample sizes $N$ , achieving low error ( $9.5$ ) even with a single training image ( $N=1$ ).

Knowl 8 — Crowd Counting Accuracy on Surveillance Video

data/table

Mean absolute errors (MAE) for pedestrian counting evaluated on the 2000-frame UCSD surveillance video benchmark across standard training/test evaluation splits:

Method	maximal	downscale	upscale	minimal	dense	sparse
Counting-by-Regression (Kong et al., 2006)	$2.07$	$2.66$	$2.78$	N/A	N/A	N/A
Counting-by-Regression (Ryan et al., 2009)	$1.80$	$2.34$	$2.52$	$4.46$	N/A	N/A
Counting-by-Segmentation (Ryan et al., 2009)	$1.53$	$1.64$	$1.84$	$1.31$	N/A	N/A
Density learning (proposed)	$1.70$	$1.28$	$1.59$	$2.02$	$1.78 \pm 0.39$	$2.06 \pm 0.59$

The four historical splits are: 'maximal' (train on frames 600:5:1400), 'downscale' (train on most crowded frames 1205:5:1600), 'upscale' (train on least crowded frames 805:5:1100), and 'minimal' (train on only 10 frames 640:80:1360). 'Dense' and 'sparse' evaluate 5-fold cross-validation chunks with 80 and 10 training frames, respectively. The proposed method outperforms direct regression baselines across all splits and performs competitively with the hybrid segmentation method, while requiring only point dot annotations rather than segmented blob annotations.

Coverage note — None omitted. All key modeling formulations, mathematical definitions, algorithms, feature extraction procedures, and experimental evaluation tables are represented in the knowls.

References

1.http://www.robots.ox.ac.uk/%7Evgg/research/counting/index.html.
2.The MOSEK optimization software. http://www.mosek.com/.
3.N. Ahuja and S. Todorovic. Extracting texels in 2.1d natural textures. ICCV, pp. 1–8, 2007.
4.S. An, P. Peursum, W. Liu, and S. Venkatesh. Efficient algorithms for subwindow search in object detection and localization. CVPR, pp. 264–271, 2009.
5.D. Anoraganingrum. Cell segmentation with median filter and mathematical morphology operation. Image Analysis and Processing, International Conference on, 0:1043, 1999.
6.O. Barinova, V. Lempitsky, and P. Kohli. On the detection of multiple object instances using Hough transforms. CVPR, 2010.
7.J. L. Bentley. Programming pearls: Algorithm design techniques. Comm. ACM, 27(9):865–871, 1984.
8.J. L. Bentley. Programming pearls: Perspective on performance. Comm. ACM, 27(11):1087–1092, 1984.
9.L. Breiman. Random forests. Machine Learning, 45(1):5–32, 2001.
10.A. B. Chan, Z.-S. J. Liang, and N. Vasconcelos. Privacy preserving crowd monitoring: Counting people without people models or tracking. CVPR, 2008.
11.S.-Y. Cho, T. W. S. Chow, and C.-T. Leung. A neural-based crowd estimation by hybrid global learning algorithm. IEEE Transactions on Systems, Man, and Cybernetics, Part B, 29(4):535–541, 1999.
12.C. Desai, D. Ramanan, and C. Fowlkes. Discriminative models for multi-class object layout. ICCV, 2009.
13.X. Descombes, R. Minlos, and E. Zhizhina. Object extraction using a stochastic birth-and-death dynamics in continuum. Journal of Mathematical Imaging and Vision, 33(3):347–359, 2009.
14.L. Dong, V. Parameswaran, V. Ramesh, and I. Zoghlami. Fast crowd segmentation using shape indexing. ICCV, pp. 1–8, 2007.
15.M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2009 (VOC2009) Results. http://pascallin.ecs.soton.ac.uk/challenges/VOC/voc2009/workshop/index.html.
16.T. Joachims, T. Finley, and C.-N. J. Yu. Cutting-plane training of structural svms. Machine Learning, 77(1):27–59, 2009.
17.D. Kong, D. Gray, and H. Tao. A viewpoint invariant approach for crowd counting. ICPR (3), pp. 1187–1190, 2006.
18.P. D. Kovesi. MATLAB and Octave functions for computer vision and image processing. School of Computer Science & Software Engineering, The University of Western Australia. Available from: http://www.csse.uwa.edu.au/∼pk/research/matlabfns/.
19.A. Lehmussola, P. Ruusuvuori, J. Selinummi, H. Huttunen, and O. Yli-Harja. Computational framework for simulating fluorescence microscope images with cell populations. IEEE Trans. Med. Imaging, 26(7):1010–1016, 2007.
20.B. Leibe, A. Leonardis, and B. Schiele. Robust object detection with interleaved categorization and segmentation. International Journal of Computer Vision, 77(1-3):259–289, 2008.
21.D. G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91–110, 2004.
22.A. N. Marana, S. A. Velastin, L. F. Costa, and R. A. Lotufo. Estimation of crowd density using image processing. Image Processing for Security Applications, pp. 1–8, 1997.
23.J. Massey, Frank J. The kolmogorov-smirnov test for goodness of fit. Journal of the American Statistical Association, 46(253):68–78, 1951.
24.F. Moosmann, B. Triggs, and F. Jurie. Fast discriminative visual codebooks using randomized clustering forests. NIPS, pp. 985–992, 2006.
25.S. K. Nath, K. Palaniappan, and F. Bunyak. Cell segmentation using coupled level sets and graph-vertex coloring. MICCAI (1), pp. 101–108, 2006.
26.T. W. Nattkemper, H. Wersing, W. Schubert, and H. Ritter. A neural network architecture for automatic segmentation of fluorescence micrographs. Neurocomputing, 48(1-4):357–367, 2002.
27.V. Rabaud and S. Belongie. Counting crowded moving objects. CVPR (1), pp. 705–711, 2006.
28.D. Ryan, S. Denman, C. Fookes, and S. Sridharan. Crowd counting using multiple local features. DICTA ’09: Proceedings of the 2009 Digital Image Computing: Techniques and Applications, pp. 81–88, 2009.
29.J. Selinummi, J. Seppala, O. Yli-Harja, and J. A. Puhakka. Software for quantification of labeled bacteria from digital microscope images by automated image analysis. Biotechniques, 39(6):859–63, 2005.
30.T. Sharp. Implementing decision trees and forests on a GPU. ECCV (4), pp. 595–608, 2008.
31.H. Tamaki and T. Tokuyama. Algorithms for the maxium subarray problem based on matrix multiplication. SODA, pp. 446–452, 1998.
32.A. Vedaldi and B. Fulkerson. VLFeat: An open and portable library of computer vision algorithms. http://www.vlfeat.org/, 2008.
33.B. Wu, R. Nevatia, and Y. Li. Segmentation of multiple, partially occluded objects by grouping, merging, assigning part detection responses. CVPR, 2008.
34.T. Zhao and R. Nevatia. Bayesian human segmentation in crowded situations. CVPR (2), pp. 459–466, 2003.

Citation

MLA

Lempitsky, V., and A. Zisserman. “Learning To Count Objects in Images”. Advances in Neural Information Processing Systems, vol. 23, 2010, https://proceedings.neurips.cc/paper_files/paper/2010/file/fe73f687e5bc5280214e0486b273a5f9-Paper.pdf.

APA

Lempitsky, V., & Zisserman, A. (2010). Learning To Count Objects in Images. Advances in Neural Information Processing Systems, 23. https://proceedings.neurips.cc/paper_files/paper/2010/file/fe73f687e5bc5280214e0486b273a5f9-Paper.pdf

Chicago

Lempitsky, V., and A. Zisserman. 2010. “Learning To Count Objects in Images”. Advances in Neural Information Processing Systems 23. https://proceedings.neurips.cc/paper_files/paper/2010/file/fe73f687e5bc5280214e0486b273a5f9-Paper.pdf.

Harvard

Lempitsky, V. and Zisserman, A. (2010) “Learning To Count Objects in Images”, Advances in Neural Information Processing Systems. Curran Associates, Inc. Available at: https://proceedings.neurips.cc/paper_files/paper/2010/file/fe73f687e5bc5280214e0486b273a5f9-Paper.pdf.

Vancouver

1. Lempitsky V, Zisserman A (2010) Learning To Count Objects in Images. Advances in Neural Information Processing Systems 23:

BibTeX

@inproceedings{lempitsky2010learning,
  title = {Learning To Count Objects in Images},
  author = {Lempitsky, Victor and Zisserman, Andrew},
  year = {2010},
  booktitle = {Advances in Neural Information Processing Systems},
  publisher = {Curran Associates, Inc.},
  volume = {23},
  url = {https://proceedings.neurips.cc/paper_files/paper/2010/file/fe73f687e5bc5280214e0486b273a5f9-Paper.pdf}
}

Metadata:DOI registry

Access the Paper

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

Open PDF

License: Authors

Learning To Count Objects in Images

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Object Counting via Linear Estimation of Spatial Density Functions

Knowl 2 — Maximum Excess over SubArrays Distance

Knowl 3 — Convex Quadratic Program for Density Estimation under MESA Distance

Knowl 4 — Cutting-Plane Optimization for MESA-Constrained Density Learning

Knowl 5 — Kernel Density Ground Truth Generation from Dot Annotations