BFD-ML-4K

Official repository: https://huggingface.co/datasets/LotusRosa/BFD-ML-4K

Companion research software: https://github.com/LotusRosa/building-facade-engineering-agent

Dataset summary

BFD-ML-4K (Building-Facade Defect Multi-Label Dataset 4K) is a frozen research corpus of 4,000 UAV images of concrete building facades. It supports full-image, multi-label defect screening with three engineering categories: hollow, spalling, and crack. Images containing none of those categories are marked no_defect.

The task is semantic screening, not object localization. No bounding boxes or segmentation masks are included. Image-level labels were produced by three civil-engineering experts and subjected to complete image-by-image cross-review.

This release is provided solely for non-commercial academic and scientific research under the custom license in LICENSE. Citation is mandatory.

Data acquisition

Field acquisition sequence used to collect UAV building-facade imagery, from ground setup and takeoff to close-range survey and the retained captured image. Prepared by the dataset authors.

Dataset contents

• Images: 4,000 privacy-sanitized JPEG files.
• Task: image-level multi-label classification.
• Defect labels: hollow, spalling, and crack.
• Negative label: no_defect, derived as the absence of all three defect labels.
• Annotation geometry: none.
• Protocol: eight mutually exclusive frozen experimental roles.
• Version: 1.0.0.

The release images preserve the original JPEG compressed image data but remove EXIF, GPS, XMP, IPTC, embedded thumbnail, camera identifier, timestamp, and JPEG comment metadata. Pixel content is not intentionally altered or recompressed.

Label semantics

Label	Meaning
hollow	Visually suspected facade hollowing observable in UAV imagery. This is not a destructively verified subsurface condition.
spalling	Visible facade material spalling.
crack	Visible facade cracking.
no_defect	None of hollow, spalling, or crack is present at image level.

Labels are non-exclusive. A defect-positive image may contain one, two, or all three defect categories.

Dataset statistics

Label	Positive images
hollow	581
spalling	1,672
crack	1,404
no_defect	1,657

There are 2,343 images with at least one defect category. Of these, 1,007 contain at least two defect categories: 700 contain exactly two, and 307 contain all three.

Frozen protocol roles

Protocol role	Images	Intended role in the frozen research experiment
core_train	2,400	Initial model construction
development_calib	200	Development-time calibration
core_safety	200	Historical-safety evaluation
adaptation_a_discovery	320	Round-A failure discovery and controlled adaptation
adaptation_a_gate	80	Independent Round-A gate
adaptation_b_discovery	320	Round-B failure discovery and controlled adaptation
adaptation_b_gate	80	Independent Round-B gate
final_test	400	Frozen final evaluation used by the associated research experiment

The final_test role is retained only to reproduce the frozen research experiment represented by this dataset release. The Building-Facade Engineering Agent product does not require a Final Test; its operational workflow supports continuous maintenance, Current Gate/Core Safety evaluation, promotion or retention, and rollback.

Frozen corpus structure

The frozen 4,000-image corpus: experimental partitions, class prevalence by partition, and exact non-overlapping image-level label combinations. Prepared by the dataset authors.

Data fields

Every split directory contains images and a neighboring metadata.jsonl. Each JSON line includes:

• file_name: image filename relative to that metadata file;
• image_id and numeric_image_id: stable pseudonymous identifiers;
• labels: list of present labels;
• hollow, spalling, crack, no_defect: binary label fields;
• protocol_role and protocol_role_label;
• capture_group and temporal_unit_id: pseudonymous acquisition-continuity group identifiers;
• label_semantics;
• width, height, mode, bytes, and release-file sha256.

metadata/all_samples.csv provides a repository-wide flat index. Summary tables, class definitions, provenance, and the machine-readable release manifest are in metadata/.

Loading the dataset

After downloading the repository, load it with Hugging Face Datasets:

from datasets import load_dataset

split_names = [
    "core_train",
    "development_calib",
    "core_safety",
    "adaptation_a_discovery",
    "adaptation_a_gate",
    "adaptation_b_discovery",
    "adaptation_b_gate",
    "final_test",
]
data_files = {name: f"data/{name}/*.jpg" for name in split_names}
dataset = load_dataset(
    "imagefolder",
    data_files=data_files,
    drop_labels=True,
)

After publication, the repository can also be loaded by its canonical Hugging Face dataset identifier.

Intended uses

Permitted uses are limited to non-commercial academic and scientific research, including:

• image-level multi-label classification research;
• research on class imbalance, calibration, continual adaptation, failure analysis, and human-governed model maintenance;
• reproducibility studies using the frozen protocol roles; and
• non-commercial comparison of research methods within the limits described here and in the license.

Out-of-scope and prohibited uses

The dataset is not designed or licensed for:

• commercial products, services, consulting, internal commercial research, or revenue-generating activity;
• operational building-safety decisions or autonomous engineering diagnosis;
• precise localization, instance counting, geometric measurement, severity assessment, or structural-condition certification;
• identifying, geolocating, or profiling people, properties, operators, or acquisition sites;
• reconstructing removed metadata; or
• redistribution, re-hosting, mirroring, or republication of the original or modified images, annotations, or subsets.

See LICENSE for the binding terms.

Relationship to UAV2K

A subset of photographs may also appear in UAV2K. BFD-ML-4K uses an independent annotation taxonomy, task definition, data organization, and modeling protocol. This repository does not include UAV2K annotations, UAV2K model weights, or UAV2K evaluation protocols. The BFD-ML-4K license governs the copies distributed in this repository and does not replace the terms governing any separately distributed UAV2K materials.

Privacy and ethical considerations

The public copies have been stripped of embedded camera, timestamp, GPS, XMP, IPTC, thumbnail, and comment metadata. Pseudonymous capture-group and temporal unit identifiers are retained solely to support leakage-aware research. Users must not attempt to recover removed metadata, infer acquisition locations, or use the data to identify people, properties, operators, or sites.

This dataset does not replace inspection by qualified professionals. Model outputs derived from it must not be treated as engineering findings without appropriate validation and human judgment.

Integrity

• checksums/SHA256SUMS.txt records SHA-256 digests for publishable files.
• audit/release_audit.json records the final local validation result.
• metadata/dataset_manifest.json records version, counts, source-manifest digest, sanitization totals, and build invariants.

License

BFD-ML-4K is released under the BFD-ML-4K Academic Research Dataset License 1.0. It permits only non-commercial academic and scientific research, requires clear citation, and prohibits redistribution or re-hosting without prior written permission. Commercial use requires a separate written license.

Dataset authors

• Yu Xia — first author and dataset repository maintainer
• Ruoyu Chen — corresponding author (chenruoyu@just.edu.cn)

Copyright in the first-party dataset release is held by Yu Xia and Ruoyu Chen.

Citation

Every public research output using or derived from BFD-ML-4K must clearly cite the dataset authors, name, version, year, and canonical repository.

@dataset{bfd_ml_4k_2026,
  author    = {Xia, Yu and Chen, Ruoyu},
  title     = {BFD-ML-4K: Building-Facade Defect Multi-Label Dataset 4K},
  year      = {2026},
  version   = {1.0.0},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/LotusRosa/BFD-ML-4K},
  note      = {Non-commercial academic and scientific research only}
}

See CITATION.cff for machine-readable citation metadata.