Flamingo Detection Using Marked Point Processes for Estimating the Size Of Populations. Détection de Flamants Roses par Processus Ponctuels Marqués pour L'estimation de la Taille des Populations

Flamingo Detection Using Marked Point Processes for Estimating the Size Of Populations

Détection de Flamants Roses par Processus Ponctuels Marqués pour L'estimation de la Taille des Populations

Stig Descamps Xavier Descombes  Arnaud Béchet   Josiane Zerubia 

Equipe-Projet Ariana, INRIA/I3S, 2004 route des Lucioles, BP 93, 06902 Sophia Antipolis, cedex (France)

Fondation Sansouire, Tour du Valat, Le Sambuc, 13200 Arles (France)

22 October 2007
30 April 2008
| Citation



In this paper, we present a new technique to automatically detect and count breeding Greater flamingos (Phoenicopterus roseus) on aerial photographs of their colonies. We consider a stochastic approach based on marked point processes also called object processes. Here, the objects represent flamingos which are defined as ellipses. The Gibbs density associated with the marked point process of ellipses is defined w.r.t the Poisson measure. Thus, the issue is reduced to an energy minimization, where the energy is composed of a regularization term (prior density), which introduces some constraints on the objects and their interactions, and a data term, which links the objects to the features to be extracted in the image. The prior energy is defined as a sum of local energies for each object. For a givenobject o, we consider the set S(o) of objects in the current configuration which overlap o. An overlapping coefficient between two objects is defined by the intersection area normalised by the minimum size between the two objects. The local energy, associated to o, is then proportional to the maximum overlapping coefficient between o and any element of S(o). The data term is also defined by a sum local energies over each object in the configuration. The local energy is obtained from the computation of a radiometric distance between pixels in the ellipse, modeling the flamingo, and pixels in the neighborhood of this ellipse.

pixels in an ellipse and pixels in the ellipse neighbordood is mapped to [−1,1]. When negative (resp. positive) this local energy favours (resp. penalizes) the object to belong to the final detection. The sign of this local energy depends on a pre-defined threshold. We propose an heuristic to locally estimate this threshold. We first estimate a gaussian law to approximate the flamingo radiometry in the color space. We compute a weighted color histogram of the image. The weight associated to each pixel is proportional to a distance computed between pixels in a disk centered on the considered pixel and its neighbors. Therefore, the contribution of the flamingos in the color histogram is enlarged. We then assume that the main mode in the histogram corresponds to flamingos. We estimate the distribution of this mode, assuming a gaussian property. The threshold is then deduced from the covariance matrix of this mode. We apply this procedure locally, on different windows and the threshhold map is obtained by interpolation. This algorithm provide a locally adaptive estimation of the threshold parameter. As a result, the only parameters required from the user are some bounding boxes for the ellipse axis. These boxes can be deduced from the usual size of flamingos and the data resolution.

Then, we sample the process to extract the configuration of objects, minimizing the energy, by a new and fast birth-anddeath dynamics, leading to the total number of birds. This sampling algorithm is embeded into a simulated annealing scheme to optimize the proposed model. A birth rate and the temperature of the simulated annealing are initialized. After each iteration, these two parameters are decreased using a geometrical law.

Each iteration is divided into three steps. The birth step consists in adding a new object on each pixel with a given probability. This probability depends on the birth rate but also on a pre-computed birth map. To compute this birth map, the data term is evaluated on each pixel for a given disk. The probability of adding a new object then depends on the data. The choice of the birth map, uniform or depending on the data, does not affect the convergence properties. Therefore, the final result is the same. But the speed of convergence is increased by using the data in the birth process. The goal of the second step is also to speed up the convergence. It consists in sorting the objects in the configuration with respect to the value of the corresponding data term. In the last step, consisting in removing, or killing, some objects in the current configuration, we first propose to kill objects having a high value for data term. For each object, taken in the data term decreasing order, we compute a death probability depending on the energy difference between the configuration and the configuration minus the considered object, and on the temperature. Note that the birth rate does not depend on the temperature, nor the configuration energy. The convergence of this algorithm has been proven in a previous work.

This approach gives counts with good precision compared to manual counts. Additionally, this approach does not need any image pre-processing or supervision w.r.t. the extraction, thus considerably reducing the overall processing time required to get the estimate. Some results are shown and compared to the results obtained by an expert. Besides, we compare the proposed approach with standard algorithm based on template matching and mathematical morphology.


Nous présentons dans cet article une nouvelle technique de détection de flamants roses sur des images aériennes. Nous considérons une approche stochastique fondée sur les processus ponctuels marqués, aussi appelés processus objets. Ici, les objets représentent les flamants, qui sont modélisés par des ellipses. La densité associée au processus ponctuel marqué d'ellipses est définie par rapport à une mesure de Poisson. Dans un cadre gibbsien, le problème se réduit à la minimisation d'une énergie, qui est constituée d'un terme de régularisation (densité a priori), qui introduit des contraintes sur les objets et leurs interactions; et un terme d'attache aux données, qui permet de localiser sur l'image les flamants à extraire. Nous échantillonnons le processus pour extraire la configuration d'objets minimisant l'énergie grâce à une nouvelle dynamique de Naissances et Morts multiples, amenant finalement à une estimation du nombre total de flamants présents sur l'image. Cette approche donne des comptes avec une bonne précision comparée aux comptes manuels. De plus, elle ne nécessite aucun traitement préalable ou intervention manuelle, ce qui réduit considérablement le temps d'obtention des comptes.


Object extraction, stochastic modeling, marked point processes, Birth and Death dynamics, environment, ecology, flamingos.

Mots clés

Extraction d'objets, modélisation stochastique, processus ponctuels marqués, dynamique de Naissance/Mort, environnement, écologie, flamants roses.

1. Introduction
2. Processus Ponctuels Marqués
3. Modèle pour l'Extraction des Flamants Roses
4. Simulation et Optimisation par Naissance et Mort
5. Résultats
6. Conclusion

BADDELEY A.J., VAN LIESHOUT M.N.M., Object recognition using Markov spatial processes, In Proceedings of International Conference Pattern Recognition, volume B, pages 136-139, 1992.

BÉCHET A., REED A., PLANTE N., GIROUX J. F., GAUTHIER G., Estimating the size of the greater snow goose population, Journal of Wildlife Management, 68(3) : pages 639-649, 2004.

DESCOMBES X., MINLOS R., ZHIZHINA E., Object Extraction Using a Stochastic Birth and Death Dynamics in Continuum, Journal of Mathematical Imaging and Vision, 33 : pages 347-359, 2009.

GREEN P.J., Reversible jump Markov chain Monte Carlo computation and Bayesian model determination, Biometrika 82, pages 711.7320, 1995.

JOHNSON A.R., CÉZILLY F., The Greater Flamingo, T & AD Poyser,2007. 

LACOSTE C., DESCOMBES X., ZERUBIA J., Point Processes for Unsupervised Line Network Extraction in Remote Sensing, IEEE Trans. Pattern Analysis and Machine Intelligence, 27(10) : pages 1568-1579, octobre 2005.

LARSEN M., Individual tree top position estimation by template voting, In Proc. of the Fourth International Airbone Remote Sensing Conference and Exhibition/21st Canadian Symposium on Remote Sensing, Ottawa, Ontario, vol. 2, pages 83-90, 1999.

VAN LIESHOUT M.C., Markov Point Processes and Their Applications, Imperial College, 2000. 

ORTNER M., DESCOMBES X., ZERUBIA J., A marked point process of rectangles and segments for automatic analysis of Digital Elevation Models, IEEE Trans. Pattern Analysis and Machine Intelligence, 30(1) : pages 105-119, 2008.

PERRIN G., DESCOMBES X., ZERUBIA J., A Marked Point Process Model for Tree Crown Extraction in Plantations, In Proceedings of International Conference on Image Processing, Genoa, 2005.

PRESTON C. J., Spatial birth-and-death processes, Bull. Internat. Statist. Inst., Vol. 46, No. 2, pp. 371-391, 1977.

RUE H., HURN M., Bayesian object identification, Biometrika, 86(3) : pages 649-660, 1999.

SOILLE J.P., Morphological Image Analysis, Springer, Berlin, 2003.

STOYAN D., KENDALL W.S., MECKE J., Stochastic Geometry and its Applications, Wiley, 1995.