Skip to content

do-me/semantic-hexbins

Repository files navigation

Semantic Hexbins

A light-weight demo app and processing scripts for geospatial semantic search. Designed for any kind of textual data with geospatial references. Repository for the paper: XXX (still to submit)

Idea

The paper describes an approach to use semantic similarity for geospatial purposes, like georeferenced social media data.

Data samples

Ranging from 8 - 32 Mb for individual posts or 0.8 - 5.1 Mb for aggreagted posts, see data folder.

Script for Data Processing & Reproduction

Scripts for data processing can be found here:

Example Queries

See the screenshots folder for query comparisons between the location-averaged and individual embedding indice.

Performance

File size

See the data directory for comparison: https://github.com/do-me/semantic-hexbins/tree/main/data

Speed

Tested devices:

  • Windows laptop with Intel i7-8550 CPU
  • Ubuntu laptop with AMD Ryzen 7 PRO 6850U
  • Android phone Samsung S9 with Exynos 9810
  • Apple iPhone 15 Pro with A17 Pro

Run times for a full layer update are significantly below 200ms with ~60ms inferencing time. Iphone 15 Pro averages around 54ms (33ms for inferencing) for 100 runs.

For comparison to a simple full-text search (GFTS) in JS see this app: https://do-me.github.io/semantic-hexbins/full_text_search_benchmark/. It benchmarks dummy data in social media style with 4 columns: lat, lon, location ID and text.

image

Screenshot results run on an M3 Max.

Previous research

About

A light-weight demo app for geospatial semantic search. Designed for text data with geospatial references.

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published