University of Illinois Urbana-Champaign

Exploiting relations among output variables for prediction and forecasting

Graber, Colin G

Content Files
GRABER-DISSERTATION-2022.pdf

Permalink

https://hdl.handle.net/2142/115463

Description

Title
Exploiting relations among output variables for prediction and forecasting
Author(s)
Graber, Colin G
Issue Date
2022-04-15
Director of Research (if dissertation) or Advisor (if thesis)
Schwing, Alexander
Doctoral Committee Chair(s)
Schwing, Alexander
Committee Member(s)
  • Forsyth, David
  • Hoiem, Derek
  • Firman, Michael
Department of Study
Computer Science
Discipline
Computer Science
Degree Granting Institution
University of Illinois at Urbana-Champaign
Degree Name
Ph.D.
Degree Level
Dissertation
Keyword(s)
  • machine learning
  • computer vision
  • structured output prediction
  • trajectory prediction
  • panoptic segmentation
  • panoptic segmentation forecasting
Abstract
In this work, we develop new approaches to model the relationships between problem variables and demonstrate that exploiting these relationships leads to improved performance for prediction and forecasting tasks. For structured output prediction, we describe a model which merges classical graphical model-based structured prediction methods with deep energy network-based approaches. We show that combining the strengths of these two approaches allows for improved performance over using them individually. Next, we introduce an approach for multi-entity trajectory prediction tasks which explicitly predicts the relationships between the entities at every point in time and uses these to select the model parameters used to forecast their future states. We show that predicting dynamic relations can lead to improved trajectory prediction performance over using a static relation graph. After this, we introduce the panoptic segmentation forecasting task and develop an initial approach to model this task. This approach functions by decomposing the scene into moving foreground components and static background components, modeling the motion of each separately. Finally, we show that introducing additional interaction modeling to the previous framework, both between all foreground instances and between foreground and background objects, leads to improved task performance and more consistent panoptic segmentation forecasts.
Graduation Semester
2022-05
Type of Resource
Thesis
Handle URL
https://hdl.handle.net/2142/115463
Copyright and License Information
Copyright 2022 Colin Graber

Owning Collections

Graduate Dissertations and Theses at Illinois PRIMARY
Graduate Theses and Dissertations at Illinois