David Paredes, Research Proficiency Presentation: 'Weakly-supervised semantic segmentation'

Thursday, September 3, 2020 - 11:30am to 1:00pm
Event Description: 

AbstractOver the past decade, deep learning has revolutionized the computer vision field. In particular, image segmentation methods have shown big improvements in typical fully-supervised learning settings. Unfortunately, these methods required datasets of costly segmentation annotations. For example, fine annotations at pixel level of an image from the CityScapes dataset required an average time of 90 minutes. Weakly-supervised segmentation methods aim to reduce the need of labor-intensive annotated datasets by training the models with weaker forms of supervision such as bounding boxes or point annotations. In this report, we explore existing approaches that try to leverage weak supervision labels for the task of semantic segmentation. Then, we propose a new approach with point annotations for semantic segmentation of immune cells in multiplex immunohistochemistry imagery. We show how our method can achieve high quality segmentation details.

Contact  for Zoom info.


Hosted By: 
Dimitris Samaras
Computed Event Type: 
Event Title: 
David Paredes, Research Proficiency Presentation: 'Weakly-supervised semantic segmentation'