Graduate Thesis Or Dissertation

 

An analysis of training methodologies for deep visual trackers Public Deposited

Downloadable Content

Download PDF
https://ir.library.oregonstate.edu/concern/graduate_thesis_or_dissertations/wp988q45c

Descriptions

Attribute NameValues
Creator
Abstract
  • This thesis considers the problem of training convolutional neural networks for online visual tracking. A major challenge for single object visual tracking is that most training sets with frame-level track annotations are quite small, due to the prohibitive cost of manual annotation. Current training approaches either supplement the annotations with other data sources (e.g., object-detection training data) or generate noisy variants of the track annotations. In either case, the data generation and training methods have ignored the fact that tracking involves sequences of decisions (one per frame) that are dependent on one another. Thus, the objectives optimized by these learning algorithms are not directly tied to the end goal of tracking performance. To further study this issue, we consider the state-of-the-art imitation learning algorithm, DAGGER, for training an online tracker. We observe that the DAGGER faces difficulty when applied to tracking, because online trackers typically experience unrecoverable failures, especially early in training. To rectify this issue we introduce, analyze, and evaluate a variation of DAGGER, called DAGGER with Resets (DAGGER), a novel imitation learning framework which maintains the theoretical properties of DAGGER and is more appropriate for training deep trackers. Our main contribution is to compare different training methods, including DAGGER and DAGGER, across a variety of datasets and multiple trackers. Our experimental results show this principled training approach and methodical random augmentation is able to outperform existing training approaches across multiple visual tracking datasets.
  • Keywords: computer science, machine learning, visual tracking, deep learning, imitation learning, computer vision
License
Resource Type
Date Available
Date Issued
Degree Level
Degree Name
Degree Field
Degree Grantor
Commencement Year
Advisor
Committee Member
Academic Affiliation
Non-Academic Affiliation
Rights Statement
Publisher
Peer Reviewed
Language
Replaces

Relationships

Parents:

This work has no parents.

In Collection:

Items