Semester

Summer

Date of Graduation

2019

Document Type

Thesis

Degree Type

MS

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Victor Fragoso

Committee Co-Chair

Powsiri Klinkhachorn

Committee Member

Powsiri Klinkhachorn

Committee Member

Nasser Nasrabadi

Abstract

Image-based localization is a crucial step in many 3D computer vision applications, e.g., self-driving cars, robotics, and augmented reality among others. Unfortunately, many image-based-localization applications require the storage of large scenes, and many camera pose estimators struggle to scale when the scene representation is large. To alleviate the aforementioned problems, many applications compress a scene representation by reducing the number of 3D points of a point cloud. The state-of-the-art compresses a scene representation by using a K-cover-based algorithm. While the state-of-the-art selects a subset of 3D points that maximizes the probability of accurately estimating the camera pose of a new image, the state-of-the-art does not guarantee an optimal compression and has parameters that are hard to tune. We propose to compress a scene representation by means of a constrained quadratic program that resembles a one-class support vector machine (SVM). Thanks to this resemblance, we derived a variant of the sequential minimal optimization, a widely adopted algorithm to train SVMs. The proposed method uses the points corresponding to the support vectors as the subset to represent a scene. Our experiments on publicly large-scale image-based localization show that our proposed approach delivers four times fewer failed localizations than that of the state-of-the-art while scaling on average two orders of magnitude more favorably.

Embargo Reason

Publication Pending

Available for download on Tuesday, August 31, 2021

Share

COinS