Learning to count with regression forest and structured labels

Following [Lempitsky and Zisserman, 2010], we seek to count objects by integrating over an object density map that is predicted from an input image. In contrast to that work, we propose to estimate the object density map by averaging over structured, namely patch-wise, predictions. Using an ensemble of randomized regression trees that use dense features as input, we obtain results that are of similar quality, at a fraction of the training time, and with low implementation effort. An open source implementation will be provided in the framework of http://ilastik.org.