Humans not only outperform current Computer Vision methods in complex general problems such as object detection/recognition, but also in domain-specific tasks such as counting cells and detecting cell divisions in time-lapse videos of mammalian embryos. This project develops a key computer vision component to reach human-like recognition performance in shape recognition for biology applications. The project provides automated methods and software tools for biology research. The project also develops a framework for addressing fundamental issues in geometry to contribute to computer vision research.
This research is rooted on pair wise symmetry and the construction of six dimensional histograms from symmetry measures. It hypothesizes that shape properties of objects can be robustly represented by marginalizing this histogram. The research team develops a method to build shape properties by performing products of marginalized histograms as to create higher level shape descriptions. The short-term goal of this project is to apply this concept to count and track overlapping cells in early mouse and human embryos. The outcome of the work includes a database of hierarchical trees of cell divisions in a mouse-embryo up to the 8-cell is essential for the analysis of particular genes in the early phases of life. The mid-term goal is to develop a shape database available to all researchers, where other computer vision methods can be tested. The long-term goal is to crack the challenging problem of understanding shapes in images. This project impacts the development of detection and recognition of objects in images and the mathematical description of shapes. It also impact on the development of a theory of leaning from big data, as it investigates shape-structures of high dimensional data.