Graduate Thesis Or Dissertation

Fine-Grained Object Recognition Under Limited Training Data

Public Deposited

Downloadable Content

Download PDF


Attribute NameValues
  • This dissertation addresses object recognition in challenging settings, where distinct object classes are visually very similar (e.g., species of birds and insects) and/or access to training examples of object classes is limited (e.g., due to the associated high costs of data annotation). In this dissertation, we present a variety of approaches aimed at images and videos, all of which are based on our hypotheses that robust fine-grained recognition requires reasoning about discriminative image or video parts and that robust learning is possible on small training sets by disentangling factors of variations in the data. Two novel search algorithms for computer vision are presented: HC-Search and HSnet Search aimed at detecting object parts and their configurations for fine-grained recognition. In addition, the problems of video summarization and human 3D pose estimation are addressed by extending Generative Adversarial Networks (GANs) with deep recurrent networks and parameterizing GAN's latent feature space with a Gaussian mixture model. Our theoretical and empirical results advance computer vision through demonstrated advantages of each approach relative to the state of the art.
Resource Type
Date Issued
Degree Level
Degree Name
Degree Field
Degree Grantor
Commencement Year
Committee Member
Academic Affiliation
Rights Statement
Embargo reason
  • Ongoing Research
Embargo date range
  • 2017-11-30 to 2018-12-29



This work has no parents.

In Collection: