Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
image
image
label
class label
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
0images
End of preview. Expand in Data Studio

BD3 crack masks

Pixel-level crack masks for 1,184 photographs of the BD3 building-defects dataset (Kottari and Arjunan, BuildSys 2024), which itself provides only image-level classes. This is the BD3 evaluation set of the SynCrack3D paper.

Set Photos BD3 class Mask
minor 487 minor crack crack pixels labelled
major 342 major crack crack pixels labelled
plain 355 normal wall all background (negatives)
total 1,184

Files

images/<id>.jpg      BD3 photograph, 512 × 512, unmodified
masks/<id>.png       crack mask: 0 background, 1 crack
polygons/<id>.json   crack polygons (LabelMe format) for every minor and major photo
manifest.json        one row per photo
SHA256SUMS

Each manifest.json row gives the photo id, its set, its near-duplicate group, the cv_fold (0–4) used in the SynCrack3D paper, the number of crack_pixels, SHA-256 hashes of the image and mask, and where the photo appears in the public BD3 releases (source).

Evaluation protocol

Near-duplicate photos share a group; never split a group between training and test. For real-only training, train on four cv_folds and score the fifth. The SynCrack3D paper reports centreline F1 at a 5 px tolerance as the main score, with precision, recall, clDice and crack IoU; a plain photo counts as a false alarm when more than 0.1% of its pixels are predicted as crack.

Licence and credit

Everything here is released under CC BY 4.0. The photographs are © the BD3 authors and are redistributed unmodified under BD3's CC BY 4.0 licence GitHub. The masks, polygons and manifest are new. Please cite when you use this dataset:

@inproceedings{bd3,
  author    = {Kottari, Praveen and Arjunan, Pandarasamy},
  title     = {BD3: Building Defects Detection Dataset for Benchmarking Computer Vision Techniques for Automated Defect Identification},
  booktitle = {Proceedings of the 11th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation},
  year      = {2024},
  pages     = {297--301},
  doi       = {10.1145/3671127.3698789}
}

TODO: Add SynCrack3D paper here.

Downloads last month
295