Paolo Andreini(DIISM, University of Siena)

April 12, 2018 – 9:30 AM

DIISM, Artificial Intelligence laboratory (room 201), Siena SI

Description

In recent years, deep learning techniques have push the state of the art in many visual recognition tasks, based on fully annotated data by human experts. Nevertheless, this annotation procedure is inherently difficult and costly, in particular in those tasks, such as semantic segmentation, which require pixel-level annotations. The aim of this seminar is to show how synthetic data generation can constitute a scalable alternative to the human ground-truth supervision in some specific domains. In particular, we developed different strategies to tackle two distinct problems, i.e. segmentation of bacterial colonies in solid agar plate images and text segmentation in natural scenes. Some preliminary results and possible future developments will be presented.