ShadowTransfer

A benchmark for measuring geographic transfer in overhead shadow detection. 4,500 human-verified shadow masks across three U.S. cities (Chicago, Miami, Phoenix) at two native NAIP resolutions (0.3 m/px, 0.6 m/px), released in two complementary forms:

• data_cities/ — raw per-city dataset organized by city, resolution, and split. Use this when you need full control over splits or want to construct your own protocols.
• data_loco/ — pre-built leave-one-city-out (LOCO) folds derived from data_cities/. Use this when you want to reproduce the paper's transfer evaluation, or when comparing a new method against the reported baselines.

Both directories contain the same underlying images and masks. data_loco/ is a re-organization of data_cities/ into the LOCO protocol with frozen, paper-matched train/val/test counts. Pick whichever matches your workflow.

---

Quick start

# Hosted at:
#   https://huggingface.co/datasets/shadow-transfer-bench/ShadowTransfer

from huggingface_hub import snapshot_download
snapshot_download(repo_id="shadow-transfer-bench/ShadowTransfer",
                  repo_type="dataset", local_dir="ShadowTransfer")

To reproduce the paper's LOCO numbers, point any segmentation training pipeline at one fold:

ShadowTransfer/data_loco/fold_0_holdout_phoenix/highres/
    train/images/    train/masks/
    val/images/      val/masks/
    test/images/     test/masks/

That's it — train/, val/, and test/ already contain the 450 / 150 / 150 images the paper uses.

---

Schema

data_cities/ — per-city raw dataset

data_cities/
├── chicago/
│   ├── highres/                          # 0.3 m/px native NAIP
│   │   ├── train/
│   │   │   ├── images/                   # 450 RGB .png, 384×384
│   │   │   ├── masks/                    # 450 binary .png (0 / 255)
│   │   │   └── masks_multiclass/         # optional, 0–6 class IDs (see below)
│   │   ├── val/
│   │   │   ├── images/                   # 150 .png
│   │   │   ├── masks/                    # 150 .png
│   │   │   └── masks_multiclass/
│   │   ├── test/
│   │   │   ├── images/                   # 150 .png
│   │   │   ├── masks/                    # 150 .png
│   │   │   └── masks_multiclass/
│   │   ├── metadata_train.json
│   │   ├── metadata_val.json
│   │   └── metadata_test.json
│   └── midres/                           # 0.6 m/px native NAIP, same layout
├── miami/                                # same layout
└── phoenix/                              # same layout

Counts (per city, per resolution): 450 train + 150 val + 150 test = 750 images. Total: 3 cities × 2 resolutions × 750 = 4,500 images.

File formats

Path	Type	Encoding
images/*.png	RGB image	8-bit, 3 channels, 384×384
masks/*.png	binary shadow mask	8-bit, 1 channel, {0, 255} (255 = shadow)
masks_multiclass/*.png	multiclass mask	8-bit, 1 channel, integer class IDs 0–6

Multiclass IDs (used in masks_multiclass/):

ID	Class
0	Background (no shadow)
1	Building / canyon shadow
2	Under-structure shadow
3	Tree-canopy dapple
4	Topography-cast shadow
5	Vehicle-cast shadow
6	Thin-linear shadow

The benchmark in the paper evaluates on binary masks only; the multiclass masks are released for downstream analysis. Image and mask filenames match within a split (images/foo.png ↔ masks/foo.png).

metadata_{split}.json — one JSON list per split, one entry per image:

{
  "original_filename":   "phoenix_session01_highres_paired_010.png",
  "random_filename":     "img_005.png",      // anonymized name on disk
  "city":                "phoenix",
  "resolution":          "highres",          // "highres" (0.3 m) | "midres" (0.6 m)
  "split":               "test",             // "train" | "val" | "test"
  "type":                "type2",            // sampling scheme tag
  "image_type":          "paired",           // "paired" if also in the other resolution
  "pair_id":             "010",              // links a paired pair across resolutions
  "center_lon":          -112.17278007840696,
  "center_lat":          33.443872697021,
  "tile_name":           "m_3311239_ne_12_030_20230917",   // source NAIP tile
  "source_session":      1,
  "annotation_session":  31,
  "session_num":         31,
  "has_annotations":     true,
  "shadow_types":        ["Building/canyon shadow",
                          "Vehicle-cast shadow",
                          "Tree-canopy dapple"]
}

The on-disk filename is random_filename. original_filename is the human-readable name. pair_id lets you join the 0.3 m/px and 0.6 m/px patches that share ground coordinates (300 paired patches per city — see paper §3.1).

---

data_loco/ — pre-built LOCO folds

Three folds, one per held-out city. Each fold contains the same train / val / test directory layout as the per-city dataset, plus a manifest.json and per-split metadata.

data_loco/
├── fold_0_holdout_phoenix/                # train: chicago + miami,  test: phoenix
│   ├── highres/                           # 0.3 m/px
│   │   ├── manifest.json                  # provenance + counts
│   │   ├── metadata_train.json
│   │   ├── metadata_val.json
│   │   ├── metadata_test.json
│   │   ├── train/
│   │   │   ├── images/                    # 450 .png (225 chicago + 225 miami)
│   │   │   ├── masks/                     # 450 .png
│   │   │   └── masks_multiclass/          # where present upstream
│   │   ├── val/
│   │   │   ├── images/                    # 150 .png (75 chicago + 75 miami)
│   │   │   ├── masks/                     # 150 .png
│   │   │   └── masks_multiclass/
│   │   └── test/
│   │       ├── images/                    # 150 .png (full phoenix test pool)
│   │       ├── masks/                     # 150 .png
│   │       └── masks_multiclass/
│   └── midres/                            # 0.6 m/px, same layout
├── fold_1_holdout_miami/                  # train: chicago + phoenix,  test: miami
└── fold_2_holdout_chicago/                # train: miami + phoenix,    test: chicago

Filename convention. In train/ and val/, files are renamed {source_city}__{original}.png (e.g. chicago__img_017.png) so the two source cities cannot collide and provenance is visible at a glance. In test/ files keep their original names because they come from a single source city. Image and mask filenames remain matched within a split.

metadata_{split}.json — same fields as the per-city metadata, plus LOCO context:

{
  // ... all per-city fields preserved as-is, plus:
  "loco_filename":         "chicago__img_017.png",
  "loco_split":            "train",         // "train" | "val" | "test" in this fold
  "loco_fold_id":          0,
  "loco_holdout_city":     "phoenix",
  "loco_resolution":       "highres",
  "source_city":           "chicago",
  "source_split":          "train",         // which per-city split it came from
  "has_masks_multiclass":  true
}

manifest.json records the build parameters, per-city counts, and the full file list — enough to re-derive the fold from data_cities/ exactly.

Counts per fold per resolution: 450 train (225 per training city) + 150 val (75 per training city) + 150 test (held-out city's full test pool).

---

Intended use

• Primary use: benchmarking shadow detection methods on overhead aerial imagery, with explicit measurement of geographic transfer (data_loco/) or of in-domain performance per city (data_cities/).
• Secondary uses: building footprint and façade extraction (binary masks act as occlusion priors); shadow-removal and de-shadowing research; domain generalization research on dense prediction tasks; pretraining for related overhead segmentation tasks; analysis of urban morphology and solar geometry from the included center_lat / center_lon and shadow type labels.

For ML transfer evaluation specifically, report results on data_loco/ using all three folds and both resolutions (six test cells per method) and use the within-city upper-bound numbers from the paper as the comparison baseline. Per-cell paired bootstrap is recommended for significance — see the paper for the exact protocol.

---

Known limitations

• Three U.S. cities only. Chicago, Miami, and Phoenix span distinct climates and morphologies but share North American grid-pattern urbanism. Generalization to dense historic European cities, informal settlements, or non-grid morphologies (e.g. Mumbai, Cairo, Marrakech) is untested.
• NAIP RGB only. No multispectral or near-infrared bands. Sensor characteristics, color processing, and acquisition conventions are NAIP-specific.
• Single fall season. All imagery comes from a single seasonal window; deciduous-canopy bare-vs-leaf-on variation is not represented.
• Native-resolution releases only. The 0.3 m and 0.6 m subsets come from separate native NAIP acquisitions, not from downsampling the same source. Do not synthesize one from the other if your goal is to study resolution transfer.
• Boundary uncertainty. Shadow edges are inherently soft; we recommend tolerant-mIoU evaluation with a ±2 px don't-care band (see paper §3.3). Strict pixel-exact metrics will systematically penalize all methods at the boundary.
• Multiclass coverage. masks_multiclass/ is provided where reliable typing was possible; sparse classes (vehicle-cast, thin-linear) have low per-image counts and are not recommended as primary evaluation targets.
• Annotation noise. Even with three-phase QC and inter-annotator-agreement monitoring, a small residual disagreement rate (≈3% of segments adjudicated as borderline) is expected.

---

License and attribution

• Source imagery (NAIP). USDA Farm Service Agency National Agriculture Imagery Program. NAIP imagery acquired through 2019 is in the U.S. public domain; later releases are published as public-domain-with-attribution by USDA-FSA APFO. Users of the imagery in derived products are asked to credit the USDA Farm Service Agency Aerial Photography Field Office (APFO).

• Annotations and metadata. The shadow masks (masks/, masks_multiclass/) and metadata files (metadata_*.json, manifest.json) are released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

• Required citation when using the dataset.

  ShadowTransfer authors. ShadowTransfer: A Geographic Transfer Benchmark
  for Overhead Shadow Detection. NeurIPS 2026 Datasets & Benchmarks Track.
  

  Please also cite USDA-FSA NAIP for the underlying imagery.

---

Hosting and DOI

• Primary host: https://huggingface.co/datasets/shadow-transfer-bench/ShadowTransfer
• DOI: assigned via the Hugging Face dataset record (visible on the dataset card).
• Mirror / archival copy: see the dataset card for the latest mirror list.

---

Documentation

Two structured-documentation artifacts accompany the release:

• DATASHEET.md — a Datasheet for Datasets in the format of Gebru et al. (2021), covering motivation, composition, collection, preprocessing, uses, distribution, and maintenance.
• Croissant metadata — machine-readable dataset description in the MLCommons Croissant format. Hugging Face auto-generates and serves this for every Hub dataset; fetch it at:

  https://huggingface.co/api/datasets/shadow-transfer-bench/ShadowTransfer/croissant
  

  This file is consumable by mlcroissant, TFDS, and any Croissant-aware loader.

---

Maintenance

Issues, errata, and corrections: file an issue on the Hugging Face dataset page or open a pull request on the accompanying GitHub repository linked from the dataset card. Versioned releases are tagged on Hugging Face; the version used for the published paper results is tagged v1.0.

For questions about the LOCO protocol or the diagnostic framework, see the paper. For questions about the annotation pipeline, see Appendix A of the paper.