Cornell University
VIEW-INVARIANT ACTION RECOGNITION IN DYNAMIC SCENES VIA SIM2REAL TRANSFER
Abstract
dc:description.abstractLearning view-invariant representation is essential to improving feature extraction for action recognition. Existing approaches cannot effectively capture details for human actions due to fast-paced gameplay and implicit view- dependent representation. In this paper, it goes beyond recognizing human actions from a fixed view and focusing on action recognition from an arbitrary view. This paper purposes a method to build an efficient data generating pipeline due to lack of original input data. This paper also provides a pipeline combining capturing modified I3D human actions features and use Multilayer Perception to achieve human action recognition and classification. The use of information captured from combination of virtual and real-life data, as well as different viewing angles, leads to high classification performance.
Degree
thesis:*- Name thesis:degree_name
- M.S., Mechanical Engineering
- Level thesis:degree_level
- Master of Science
- Discipline thesis:degree_discipline
- Mechanical Engineering
- Grantor
- Cornell University
- Year dc:date.issued
- 2023
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Wang, Yuhan
- Chair dc:contributor.chair
-
- Ferrari, Silvia
- Committee member dc:contributor.committeemember
-
- Hariharan, Bharath
Subjects
dc:subject × 3Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Dc Identifier Bibid
- 16219432
- Dc Identifier Other
-
ProQuest Submission ID: 11820
ProQuest Publication ID: 30528529
ProQuest Submission ID: 11839
ProQuest Publication ID: 30568032