Events Calendar

PhD Preliminary Exam – Khoi Nguyen

Few-shot Segmentation

This research report addresses the problem of few-shot segmentation. The goal of segmentation is to label every pixel of the image with a semantic or instance class (different objects of the same class will have different ids). Few-shot segmentation means that in the testing, we are asked to segment query images with completely different object classes from the training set with few support examples and their ground-truth labels. We focus on two types of segmentation: (1) semantic segmentation, where ground-truth support segments are semantic level and we need to do the semantic segmentation for query images; (2) instance segmentation, where ground-truth support instance segmentation are given and thus we are asked to do instance segmentation for query images. To address these problems, we make three hypotheses. First, we believe that more discriminative (high activation for a specific class only) and ensemble of features should be extracted from support ima! ges to represent novel classes such that they are not over-fitted to these support images. To this end, we use feature weighting and feature boosting to extract features from support images. Second, since the training and testing classes are disjoint, we hypothesize that generative (i.e GAN) and self-supervised models learn move general features than discriminative models. In particular, discriminative models focus on learning features that are well-described the training classes and ignore others. Third, in few-shot setting, we make a hypothesis that object boundary and object shape are also very important cues along with common feature similarity. To achieve this, we make two contributions: (1) formulate an Mumford-Shah like optimization problem and (2) predict normalized object shapes at every location belongs to the object and transformation matrices to transform them into object masks. Our preliminary experimental evaluation demonstrates the advantages of our hypothese! s relative to existing work. In this report, we also lay out o! ur futur e evaluation plan and publication goals toward a Ph.D. thesis.

Major Advisor: Sinisa Todorovic
Committee: Fuxin Li
Committee: Alan Fern
Committee: Mike Rosulek
GCR: Jay Kim

Thursday, March 12 at 10:00am to 12:00pm

Kelley Engineering Center, 1005
110 SW Park Terrace, Corvallis, OR 97331

Event Type

Lecture or Presentation

Event Topic

Research

Organization
Electrical Engineering and Computer Science
Contact Name

Calvin Hughes

Contact Email

calvin.hughes@oregonstate.edu

Subscribe
Google Calendar iCal Outlook

Recent Activity