Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Application of Disentanglement to Map Registration Problem

Hae Jin Song, Patrycja Krawczuk, Po-Hsuan Huang

TL;DR

This paper hypothesizes that a combination of $\beta$-VAE-like architecture and adversarial training will achieve both the disentanglement of the geographic information and artistic styles and generation of new map tiles by composing the encoded geographic information with any artistic style.

Abstract

Geospatial data come from various sources, such as satellites, aircraft, and LiDAR. The variability of the source is not limited to the types of data acquisition techniques, as we have maps from different time periods. To incorporate these data for a coherent analysis, it is essential to first align different "styles" of geospatial data to its matching images that point to the same location on the surface of the Earth. In this paper, we approach the image registration as a two-step process of (1) extracting geospatial contents invariant to visual (and any other non-content-related) information, and (2) matching the data based on such (purely) geospatial contents. We hypothesize that a combination of $β$-VAE-like architecture [2] and adversarial training will achieve both the disentanglement of the geographic information and artistic styles and generation of new map tiles by composing the encoded geographic information with any artistic style.

Application of Disentanglement to Map Registration Problem

TL;DR

This paper hypothesizes that a combination of -VAE-like architecture and adversarial training will achieve both the disentanglement of the geographic information and artistic styles and generation of new map tiles by composing the encoded geographic information with any artistic style.

Abstract

Geospatial data come from various sources, such as satellites, aircraft, and LiDAR. The variability of the source is not limited to the types of data acquisition techniques, as we have maps from different time periods. To incorporate these data for a coherent analysis, it is essential to first align different "styles" of geospatial data to its matching images that point to the same location on the surface of the Earth. In this paper, we approach the image registration as a two-step process of (1) extracting geospatial contents invariant to visual (and any other non-content-related) information, and (2) matching the data based on such (purely) geospatial contents. We hypothesize that a combination of -VAE-like architecture [2] and adversarial training will achieve both the disentanglement of the geographic information and artistic styles and generation of new map tiles by composing the encoded geographic information with any artistic style.
Paper Structure (17 sections, 3 equations, 8 figures, 1 table)

This paper contains 17 sections, 3 equations, 8 figures, 1 table.

Table of Contents

  1. Motivation and Hypothesis
  2. Motivation
  3. Hypothesis
  4. $\beta$- VAE with Adversarial Training
  5. Background and Related Work
  6. Datasets
  7. Digits
  8. Maps: Stamen Dataset
  9. Architectures and Cost Functions
  10. Baseline: Domain Specific Encoders-Decoders Networks
  11. $\beta$-VAE with Friend and Enemy Networks (FENs)
  12. Experiments
  13. Baseline: Domain Specific Encoders-Decoders Networks
  14. $\beta$-VAE with Friend and Enemy Networks (FENs)
  15. Training Results
  16. ...and 2 more sections

Figures (8)

  • Figure 1: Views of the same geographical location (Paris, France). Left: OpenStreetMap. Right: Google Satellite Image.
  • Figure 2: Digits datasets.
  • Figure 3: The same geographic area shown in 3 different styles : Terrain, Watercolor and Toner.
  • Figure 4: The image with the style label A passes only through the Encoder with the label A. The representation produced by the Encoder A is fed to all of the decoders in the model (Decoders A,B and C). The reconstruction loss is calculated by comparing the output of each decoder to the ground truth images for each of the style (A,B and C). The whole model consists of 3 encoders and 3 decoders.
  • Figure 5: Overview of our $\beta$-VAE with a Friend and an Enemy Networks.
  • ...and 3 more figures