1. Introduction
Parsing images of biological substances has become one of the important applications of computer vision [8]. In many biologically-important scenarios it is necessary to deal with images of overlapping objects or organisms. In recent years, several approaches have been proposed that can parse images when objects have simple blob-type shapes (e.g. cell cultures) [1], [2], [4], [7], [23]. Less attention, however, has been paid to images containing more complex organisms exhibiting significant shape and pose variations, such as worms, larvae, and bacilli. The sheer importance of such model organisms for biomedical studies calls for further improvement of parsing approaches for this class of images.
We consider the image parsing tasks for three different organisms that are popular in biomedical research. In each case, parsing is hard because of a certain shape variability of individual organisms as well as organism overlap and crowding. Although the three organisms are very different biologically, we approach the corresponding parsing tasks with a unified framework that first uses a specially-designed deep network to propose hypotheses about the shapes of individual organisms and then use integer programming to pick a viable hypotheses set.