sukiewang/poi-benchmark

POI Benchmark: Multi-City Multimodal Points of Interest

A large-scale multimodal benchmark pairing Points of Interest (POIs) with street-view imagery, aerial grid photos, and satellite imagery across 10 major cities on 3 continents.

Cities

Beijing, Chengdu, Guangzhou, Hong Kong, Shanghai, Shenzhen, London, Melbourne, New York, Sydney.

Contents

Path	Size	Type	Description
metadata_aligned.tar	8.6 GB	11 JSON files	Enriched & aligned POI metadata per city (one JSON per city + index)
poi_data.tar	30 GB	JSONL / JSON	Raw POI text records and evaluation queries
grid_images/<city>.tar	~111 GB total	JPG	Shop-front / storefront grid images
streetview_grids/<city>.tar	~402 GB total	JPG	Street-view panorama grids
streetview_sampled/<city>.tar	~1.4 TB total	JPG	Higher-resolution street-view grids (~500m sampling)
merged_satellite/<city>.tar	~1.5 TB total	PNG	Merged multi-tile satellite images

Total: ~3.4 TB across ~2.6M images + structured POI records.

Archive format

All image collections are packaged as uncompressed .tar archives (one per city per modality). Images are already JPG/PNG — additional compression would not reduce size.

Each tar contains a top-level directory matching the city name, e.g. inside grid_images/beijing.tar:

beijing/
  beijing_0_grid.jpg
  beijing_1_grid.jpg
  ...

Load with datasets

from huggingface_hub import hf_hub_download
import tarfile, json

# 1. Metadata
path = hf_hub_download(
    repo_id="sukiewang/poi-benchmark",
    filename="metadata_aligned.tar",
    repo_type="dataset",
)
with tarfile.open(path) as t:
    t.extractall("./metadata")

# 2. One city's grid images
path = hf_hub_download(
    repo_id="sukiewang/poi-benchmark",
    filename="grid_images/beijing.tar",
    repo_type="dataset",
)

Or stream directly with WebDataset:

import webdataset as wds
url = "https://huggingface.co/datasets/sukiewang/poi-benchmark/resolve/main/grid_images/beijing.tar"
ds = wds.WebDataset(url).decode("pil")
for sample in ds:
    print(sample["__key__"], sample["jpg"].size)
    break

Intended use

• POI recognition / retrieval from visual context
• Vision-language grounding (match POI text to streetview / satellite)
• Urban scene understanding
• Cross-city generalization benchmarks

Citation

@misc{wang2026poibench,
  title  = {Multi-City POI Multimodal Benchmark},
  author = {Wang, Sukie},
  year   = {2026},
  url    = {https://huggingface.co/datasets/sukiewang/poi-benchmark},
}

License

Released under CC-BY-4.0. Third-party imagery (satellite tiles, street-view) remains under the original providers' terms; this release aggregates data for research use only.