Abstract
dc:description.abstractIn this work, we analyze and apply various recent techniques in visual attribute recognition and labeling on a common benchmark dataset in order to motivate the design of a novel framework for this task. Using the large scale COCO Attributes dataset as our benchmark, we systematically investigate recent techniques and advances in the attribute recognition literature in a unified fashion, drawing comparisons and insights from the results. We leverage these insights to propose a new models and loss function to better model the function space of attribute prediction models. Our proposed techniques are based on standard efficient building blocks readily available to researchers and practitioners, are conceptually simple, and are theoretically grounded, while giving state-of-the-art results, and generalises to various sub-domains of attribute recognition. Experiments and ablation studies performed on our model and other methods further corroborate the design decisions for our framework and shed light on possible future avenues of investigation. Our hope is that our model serves as a tool to embed strong visual attribute recognition in more complex visual reasoning tasks and pipelines.
Degree
thesis:*- Level thesis:degree_level
- Masters
- Department dc:contributor.department
- Computer Science
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2019
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Agrawal, Varun
- Advisor dc:contributor.advisor
-
- Hays, James
- Committee members dc:contributor.committeemember
-
- Essa, Irfan
- Parikh, Devi
Subjects
dc:subject × 6Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- http://hdl.handle.net/1853/61813