Emmanuel Maggiori: Optimizing Partition Trees for Multi-Class Segmentation with Shape Prior

Abstract: Image segmention is the task of partitioning an image into a non-overlapping set of regions. It is well known that the inclusion of high level cues (e.g., object shape) can significantly enhance the performance of the techniques. However, their inclusion in typical segmentation schemes is not trivial and currently constitutes an active research area. In our work we posed the problem of segmentation with shape priors as finding minimal cuts on binary partition trees (BPTs). These trees are hierarchical image partitions, usually computed in a bottom-up manner, then pruned to detect objects. Contrary to the typical static use of BPTs, we proposed an optimization algorithm that prunes and regrafts tree branches so that the structure can better represent the underlying objects, taking shape priors into account. Theoretical guarantees help reducing the search space and make the optimization efficient. Our experiments show that the approach succeeds in incorporating shape information into multi-label segmentation, outperforming the state-of-the-art.

Bio (Emmanuel Maggiori): I completed my Engineering degree in 2014 at UNCPBA, Argentina. In the same year, I joined Inria as an intern for six months in Ayin and Stars teams, supervised by Yuliya Tarabalka and Guillaume Charpiat. In January 2015 I joined Titane team to pursue a PhD at the same institution.