sdan/geospot-unified

GeoSpot WebDataset (staged)

This repository hosts WebDataset .tar shards for streaming training.

Layout

• StreetView (all in one logical set): data/streetview/**
  • data/streetview/sv2/ – pano .json + .000/.090/.180/.270.jpg (JSON uses lng)
  • data/streetview/us_local_s128/ – single-view .jpg + .json
  • data/streetview/sv1/<country>/ – per-heading samples (panoid_000.jpg + panoid_000.json)
• Everything else: data/other/**
  • data/other/mp16pro/, data/other/dress_s128/ – single images + JSON
  • data/other/osv5m/ – dashcam images + JSON
  • data/other/msls/ – Mapillary sequences (single images + JSON)
  • data/other/cvcities/ – ground pano .jpg + .sat bytes + JSON

Manifests

Note: dress and us_local were resharded into ~128MiB parts for faster streaming.\n Canonical manifests (manifests/train.txt, etc.) now point at dress_s128/ and us_local_s128/.\n Shard lists are in manifests/*.txt (paths inside this repo). For streaming, prefix with the HF resolve URL:

repo = "sdan/geospot-unified"
base = f"https://huggingface.co/datasets/{repo}/resolve/main/"
paths = open("manifests/train.txt").read().splitlines()
urls = [base + p for p in paths]

Quick starts:

• StreetView only: use manifests/streetview.txt
• Everything else: use manifests/other.txt

Streaming (WebDataset)

import webdataset as wds

ds = (
    wds.WebDataset(urls)
      .decode("pil")
      .to_tuple("jpg", "json")
)

Indexing / Segmentation

• metadata/shards.parquet (or metadata/shards.jsonl): shard-level index (domain/group/path/size/split).
• metadata/shards_enriched.parquet + manifests/geo/**: optional geo enrichment + region/country shard lists.
  • Generate/update these with scripts/hf_enrich_wds_metadata.py after the main shard upload completes.

Note: train/validation/test manifests are shard-level hashes, not spatial splits. For a real geolocation holdout split, compute a deterministic split from (lat, lon) at training time.