cif-dataset

# Cracks in the Foundation A civil-infrastructure visual inspection dataset for instance segmentation with 6 defect/condition categories: **Algae** · **Crack** · **Net-Crack** · **Crack with Precipitation** · **Rust** · **Spalling** Each sample is either a full-resolution inspection image or a 1024×1024 tile derived from one. Tiled samples carry extra fields (`tile_row`, `tile_col`, `file_name_original`, …) that are `None` for full-resolution samples. --- ## Splits Each split is its own parquet shard *and* its own dataset config. `load_dataset(repo)` returns all six in a `DatasetDict`. Naming a config — `load_dataset(repo, "train_tiled", split="train")` — is a **true selective download**, fetching only that shard. | Split | Contents | |---|---| | `train_full` | full-resolution training images | | `val_full` | full-resolution validation images | | `test_full` | full-resolution test images | | `train_tiled` | 1024×1024 tiles, training | | `val_tiled` | 1024×1024 tiles, validation | | `test_tiled` | 1024×1024 tiles, test | --- ## Load ```python from datasets import load_dataset # Load all six splits at once (single DatasetDict): all_splits = load_dataset("ibm-research/cif-dataset") # Selective: download only the tiled training shard. # Each split is also exposed as its own config — naming a config # downloads only its parquet files. ds = load_dataset("ibm-research/cif-dataset", "train_tiled", split="train") ``` --- ## Schema Every sample has the same fields regardless of split: ```python sample = ds[0] sample["image_id"] # int — unique image identifier sample["image"] # PIL.Image sample["file_name"] # str — original filename sample["width"] # int — image width in pixels sample["height"] # int — image height in pixels # Tiled-only fields (None for full-resolution samples): sample["tile_row"] # int | None — top-left row of the tile in the original image sample["tile_col"] # int | None — top-left column sample["file_name_original"] # str | None — filename of the parent image sample["width_original"] # int | None — parent image width sample["height_original"] # int | None — parent image height # Annotations (COCO convention): obj = sample["objects"] obj["id"] # List[int] obj["category_id"] # List[int] — 1=Algae 2=Crack 3=Crack(net) 4=Crack+precip 5=Rust 6=Spalling obj["bbox"] # List[[x, y, w, h]] — pixels, COCO origin (top-left) obj["area"] # List[float] obj["iscrowd"] # List[int] obj["segmentation"] # List[List[List[float]]] — polygons as flat [x1,y1,x2,y2,...] lists ``` Distinguish sample type at runtime: ```python is_tile = sample["tile_row"] is not None ``` --- ## Visualize ```bash pip install datasets fiftyone ``` ```python import tempfile from pathlib import Path import fiftyone as fo from datasets import load_dataset CATS = {1: "Algae", 2: "Crack", 3: "Crack (net-crack)", 4: "Crack with precipitation", 5: "Rust", 6: "Spalling"} ds = load_dataset("ibm-research/cif-dataset", split="test_full") tmp = Path(tempfile.mkdtemp()) fo_ds = fo.Dataset("cif_test_full", overwrite=True) for s in ds: img_path = tmp / Path(s["file_name"]).name s["image"].save(img_path) W, H = s["width"], s["height"] dets, polys = [], [] obj = s["objects"] for i, cid in enumerate(obj["category_id"]): label = CATS.get(cid, str(cid)) x, y, w, h = obj["bbox"][i] dets.append(fo.Detection(label=label, bounding_box=[x/W, y/H, w/W, h/H])) for poly in obj["segmentation"][i]: if len(poly) < 6: continue pts = [[poly[j]/W, poly[j+1]/H] for j in range(0, len(poly), 2)] polys.append(fo.Polyline(label=label, points=[pts], filled=True, closed=True)) fo_ds.add_sample(fo.Sample( filepath=str(img_path), detections=fo.Detections(detections=dets), segmentations=fo.Polylines(polylines=polys), )) session = fo.launch_app(fo_ds) session.wait() ``` Opens the FiftyOne app at `http://localhost:5151` with bounding boxes and segmentation overlays. --- ## Acknowledgment We would like to sincerely thank Finn Bormlund and Svend Gjerding (Sund & Baelt), Jens Häggström (Trafikverket), Raphael von Thiessen (Innovation-Sandbox for AI, Office for Economy, Kanton Zürich), and the Dübendorf Air Base for granting us the opportunity to collect, analyze, and disseminate the images and defect data included in this publication. --- ## Citation ```bibtex @dataset{cracks_in_the_foundation, author = {}, title = {Cracks in the Foundation}, year = {2025}, publisher = {HuggingFace}, url = {https://huggingface.co/datasets/ibm-research/cif-dataset}, } ```