Victor is a lab leader at Samsung AI Center, Moscow, and an associate professor at Skoltech, where he is leading the computer Vision Group.
He will give a keynote on:
New Single-Stage Approaches to Instance Segmentation
Instance segmentation is a key computer vision task for plant phenotyping. Two-stage systems that perform instance segmentation using object detection followed by object mask estimation (such as Mask-RCNN) have been hugely successful. At the same time, there is an interest in the community in single-stage instance segmentation methods that solve the same task by labeling individual pixels akin to semantic segmentation. Such single-stage approaches are much simpler and more computationally efficient. Yet they need to solve an intricate problem of choosing the pixel label set that can successfully encode the instance segmentation task while being amenable for training. In this talk, I will discuss two new single-stage approaches to instance segmentation that use two different label spaces and loss functions, and both achieve competitive results on several tasks including leave segmentation tasks in plant phenotyping. This is joint work with Victor Kulikov and Victor Yurchenko.