LSAA-12K / README.md
yisui's picture
Release LSAA-12K train and test archives with dataset card and checksums
474d548 verified
|
Raw History Blame Contribute Delete
4.01 kB
metadata
license: cc-by-nc-sa-4.0
language:
  - en
task_categories:
  - image-segmentation
  - text-to-image
tags:
  - facade
  - architecture
  - semantic-segmentation
  - sage-facade
size_categories:
  - 10K<n<100K

LSAA-12K

LSAA-12K contains 12,097 building facade samples with paired images, semantic label maps, JSON annotations, and English text prompts. It is released for the SAGE-Façade project, the implementation of SAGE-Façade: Semantic-Aligned Guidance for Exemplar-Based Façade Synthesis.

Splits

Split Samples Archive
train 11,795 train.zip
test 302 test.zip
Total 12,097

The archives preserve the original split and directory layout. No source files are modified, removed, or re-encoded. Each sample has four files with the same filename stem:

train/                        test/
  images/<sample_id>.jpg         images/<sample_id>.jpg
  labels/<sample_id>.png         labels/<sample_id>.png
  jsons/<sample_id>.json         jsons/<sample_id>.json
  prompts/<sample_id>.txt        prompts/<sample_id>.txt
  • images: facade images in JPEG format.
  • labels: semantic label maps in PNG format. Keep their original values/colors; do not treat them as ordinary RGB photographs or apply lossy compression.
  • jsons: original JSON annotations. Inspected examples use LabelMe-style fields such as shapes, label, points, and shape_type.
  • prompts: English descriptions, for example A building facade with 6 floors and 50 windows.

Sample IDs match across all four directories in each split. No sample ID is shared between train and test. This filename check does not establish absence of visually similar images or related building views.

Download and extract

Copy this dataset repository's ID (yisui/LSAA-12K) from the page URL, then substitute it in the command below:

python -m pip install -U huggingface_hub
hf download yisui/LSAA-12K train.zip test.zip SHA256SUMS \
  --repo-type dataset --local-dir ./LSAA-12K
cd LSAA-12K
sha256sum -c SHA256SUMS
unzip train.zip
unzip test.zip

To download only the test split, request test.zip and extract that archive. Public, ungated datasets can be downloaded without logging in. For private or gated access, log in with hf auth login first and obtain access as needed.

These ZIP files are an archival distribution for the project's existing file layout. Use hf download or huggingface_hub.snapshot_download and extract them before training; this release does not define a structured datasets.load_dataset table or guarantee a Dataset Viewer preview.

SHA256SUMS provides checksums of both archives. dataset_stats.json records sample counts, file counts, source sizes, archive sizes, and packaging checks. ZIP member paths, uncompressed sizes, and CRCs were checked after packaging.

Usage and provenance

Use the train split for model development and the test split for evaluation. Refer to the project's dataset documentation for the expected layout and label conventions. This card does not infer class-index or color mappings from annotation examples.

The project README declares LSAA-12K a derivative work based on LSAA under CC BY-NC-SA 4.0. The metadata above follows that declaration. Attribute the LSAA source dataset and the SAGE-Façade/LSAA-12K creators when reusing the data, preserve relevant source notices, and follow the CC BY-NC-SA 4.0 terms. The project's code license (Apache-2.0) is separate from the dataset license.

Citation

Please cite the SAGE-Façade paper and the original LSAA dataset when using this release. An official BibTeX entry should be added when supplied by the authors; this card does not invent publication metadata.