Résumé
We present a method combining marked point processes and convolutional neural networks applied to the detection of small objects in optical satellite images. In such images, objects are densely scattered, and visual information is scarce. The point process framework allows factoring in priors to account for object interactions. Classical point process approaches make use of contrast measures to account for object location. These fail when contrast is low and visual aspect is varied. We replace those with terms build from convolutional neural network outputs. Moreover, we propose a method to learn the parameters of the point process energy model. We show our approach improves results from the straight convolutional neural network outputs. The code will be available at github.com/Ayana-Inria/