A seminar of the A3SI team of the LIGM (mixed research unit of Paris Est University) will take place on Monday, July 2nd from 1:30 pm to 2:30 pm, room 3005 (ESIEE PARIS).
Abstract: In a first part, I will sketch how to realize an end-to-end learning of a segmentation pipeline involving a watershed computation [Wolf et al., "Learned Watershed", ICCV 2017].
In a second part, I will discuss ongoing work on instance segmentation, which can be cast as a graph partitioning problem.
The majority of models developed in this context have relied on purely attractive interactions between graph nodes. To obtain more than a single cluster, it is then necessary to manually set a merging or splitting threshold, or to pre-specify a desired number of clusters.
A notable exception to the above is multicut partitioning / correlation clustering, which allows for repulsive in addition to attractive interactions, and which automatically determines an optimal number of clusters. Unfortunately, the multicut problem is NP-hard.
In response, we propose an objective function that allows for both repulsive and attractive interactions, but which we show can be solved to optimality by a greedy algorithm. At the time of writing, the new scheme gives the state of the art results on the ISBI connectomics challenge.
Joint work with Steffen Wolf, Constantin Pape, Nasim Rahaman, Alberto Bailoni, Anna Kreshuk, Ullrich Koethe.