GeoSEAN11K

GeoSEAN11K is a route-balanced street-level image dataset for country-level geolocation across the 11 ASEAN countries. Every image is labelled both by country and by whether readable scene text is present, so text-based and scene-based geolocation evidence can be studied separately rather than mixed into a single benchmark.

The dataset contains 11,000 images, balanced to exactly 500 text-bearing and 500 no-text images per country. It extends the 4,850-image ASEAN Geolocation dataset (GeoSEAN) over the same 11 country classes.

Dataset summary

• Task: country-level image classification / image geolocation
• Countries: 11 ASEAN countries
• Images: 11,000 (1,000 per country)
• Routes: 5,500 text-bearing, 5,500 no-text
• Labels: country code, text presence, collection source
• Sources: 8,495 Google Street View panoramas, 793 Google Images, 1,712 GeoGuessr-sourced street-level images
• Image format: crawled Street View at 640x640, GeoGuessr at 1536x662, other sources at original resolution

The two routes

Street-level images carry two distinct kinds of geographic evidence: readable text (signs, shop names, road markers) and visual scene characteristics (architecture, vegetation, road layout). These benefit from different processing approaches, so every image is assigned to one of two routes:

• text — readable scene text is present. Suited to OCR and vision-language models.
• no_text — no readable scene text. Must be classified from scene appearance alone.

Route membership is fixed by manual curation, not by an automatic text detector, and is balanced at 500 images per route per country. This makes the two routes directly comparable across countries without reweighting.

Dataset composition

Per country, route, and collection source. SV: Street View; GI: Google Images; GG: GeoGuessr-sourced street-level imagery.

Country		Text route			No-text route		Total
	SV	GI	GG	SV	GI	GG
Brunei (BN)	410	90	0	500	0	0	1,000
Cambodia (KH)	399	86	15	400	1	99	1,000
Indonesia (ID)	295	127	78	316	0	184	1,000
Laos (LA)	455	28	17	458	2	40	1,000
Malaysia (MY)	377	73	50	226	0	274	1,000
Myanmar (MM)	404	96	0	496	2	2	1,000
Philippines (PH)	332	99	69	363	1	136	1,000
Singapore (SG)	379	39	82	223	0	277	1,000
Thailand (TH)	371	66	63	180	0	320	1,000
Timor-Leste (TL)	420	80	0	499	1	0	1,000
Vietnam (VN)	499	1	0	493	1	6	1,000
Total	4,341	785	374	4,154	8	1,338	11,000

Data fields

The metadata file is metadata.csv at the repository root:

Column	Type	Description
file_name	string	Relative path to the image, e.g. images/BN/BN_text_047_b18e1f8d2cf2.jpg.
image_id	string	Stable per-image identifier.
country_code	string	Country label: BN, KH, ID, LA, MY, MM, PH, SG, TH, TL, VN.
text_presence	string	Route label: text or no_text.
selection_rank	integer	Curation rank within the country and route (1-500).
content_sha256	string	SHA-256 of the image bytes.
source	string	Collection source: street_view, google_images, or geoguessr.

Public filenames encode the country, the route, the selection rank, and the first 12 hex characters of content_sha256. They deliberately carry no source identifier, so a model cannot shortcut to the collection source from the filename.

Verifying integrity

Because content_sha256 covers the image bytes, the download can be verified without a reference copy:

import hashlib
import pandas as pd
from datasets import load_dataset

meta = pd.read_csv("metadata.csv")
bad = [r.file_name for r in meta.itertuples()
       if hashlib.sha256(open(r.file_name, "rb").read()).hexdigest() != r.content_sha256]
print(f"{len(bad)} mismatched files")

Directory structure

README.md
metadata.csv
images/
  BN/
    BN_no_text_001_6a56f41292a8.jpg
    BN_text_047_b18e1f8d2cf2.jpg
    ...
  ID/
  KH/
  LA/
  MM/
  MY/
  PH/
  SG/
  TH/
  TL/
  VN/

Usage

from datasets import load_dataset

ds = load_dataset("0xRafie/geosean11k", split="train")

# Study the two routes separately
text_route = ds.filter(lambda x: x["text_presence"] == "text")
no_text_route = ds.filter(lambda x: x["text_presence"] == "no_text")

Collection method

Street View images were collected from panoramas discovered by a breadth-first search crawl of the road network. Every crawled panorama was tested against the national boundary polygon of the target country and kept only when its coordinates fell inside, so Street View country labels follow from panorama coordinates rather than from visual judgment. Google Images and GeoGuessr-sourced images were inherited from the GeoSEAN dataset.

Intended uses

• Country-level image geolocation for the ASEAN region.
• Separating scene-text evidence from scene-appearance evidence in geolocation models.
• Benchmarking OCR and vision-language models on street-level signage.
• Explainable geolocation research and regional representation studies.

Out-of-scope uses

• Identifying the exact location of an image; only country labels are provided.
• Surveillance, person identification, or face recognition.
• Any use that attempts to re-identify individuals, vehicles, or private property appearing incidentally in street-level imagery.

Limitations

• Country balance is by construction, not by natural prevalence; accuracy on this dataset does not transfer directly to unbalanced real-world distributions.
• Source composition varies substantially across countries and routes. Brunei and Timor-Leste have no GeoGuessr images at all, while over a quarter of the Thai no-text route is GeoGuessr-sourced. Per-country results can therefore reflect source characteristics as well as country characteristics.
• Route assignment reflects human judgment of whether text is readable, which is a graded property rather than a binary one.
• Street View coverage is uneven across the region and is denser along major roads, so images skew toward roadside scenes.

Licensing and provenance

All imagery was obtained from publicly accessible sources and is redistributed solely for non-commercial research use. The dataset is not licensed for commercial use. Source imagery remains subject to the terms of its original providers; users are responsible for compliance with those terms in their own jurisdiction and use case.

Related dataset

• 0xRafie/asean-geolocation — GeoSEAN, the 4,850-image predecessor over the same 11 country classes.