Event
Ph.D. Research Proposal Exam: Chinmaya Devaraj
Monday, May 3, 2021
3:30 p.m.
https://umd.zoom.us/j/4702285682?pwd=NEZyQ0ord1BZeWpSUFMzYWcxaFduQT09
Maria Hoo
301 405 3681
mch@umd.edu
ANNOUNCEMENT: Ph.D. Research Proposal Exam
Name: Chinmaya Devaraj
Committee:
Professor Yiannis Aloimonos (Chair)
Professor Behtash Babadi
Professor Dinesh Manocha
Date/time: Monday, May 3rd 2021 at 3:30 pm to 5:30 pm
Location: https://umd.zoom.us/j/4702285682?pwd=NEZyQ0ord1BZeWpSUFMzYWcxaFduQT09
Title: Revisiting Symbolic Approaches by Introducing Meta-verbs For Action Understanding
Abstract: Research in action recognition in the last few years is often looked at as obtaining the best performing model in various datasets, ignoring the constituents of actions and the goal of the action. We look beyond the traditional paradigm of supervised learning to understand the components of actions and substantially use the goal of action to develop methods for the zero-shot recognition of actions. In this thesis, we mainly limit our attention to manipulation actions, although the ideas can be extended to whole-body actions without many modifications. The core idea in our work is to obtain mid-level symbolic representations, which we refer to as primitives of manipulation actions that have a definite meaning in different domains. We develop methods in obtaining these mid-level symbolic representations like object motion, object-state changes, tools & objects involved, etc. We use this mid-level representation along with deep features from recent action recognition architectures to perform zero-shot learning, transfer learning.
