--- license: cc-by-sa-4.0 pretty_name: 'MCF-Endface: Annotated 4-core multi-core fiber end-face images' language: - en tags: - mcf - splicer - optical-fiber - endface size_categories: - n<1K configs: - config_name: default data_files: - split: train path: images/** --- # MCF-Endface End-face images of **4-core multi-core fibers (MCF)** captured on a fusion splicer, annotated with the position and radius of each core and the position of the orientation marker. The dataset is intended for training and evaluating models that locate cores and markers, which is the input needed to rotationally align two MCFs before splicing. ## At a glance | | | |---|---| | Images | 216 (108 X/Y pairs) | | Image size | 806 × 802 px, cropped and centred on the fiber cladding | | Cores | 864 (4 per image) | | Markers | 216 (1 per image) | | Fiber types | Normal: 209 · Trenched: 7 | | Capture period | 2026-03-05 → 2026-10-16 (27 capture days) | | Split | `train` only | ## Repository layout ``` annotations.json # all annotations, COCO-style (single file) images/ ├── metadata.jsonl # same annotations, one line per image (used by load_dataset) ├── 000001.jpg ├── 000002.jpg └── ... MCFDataset.py # small loader for annotations.json + images (no `datasets` needed) ``` `annotations.json` and `images/metadata.jsonl` contain the same annotations in two layouts. ## Annotation fields Each image has one record: | Field | Type | Description | |---|---|---| | `id` | int | Unique image id (images are numbered in capture order) | | `file_name` | string | Image file in `images/`, e.g. `000001.jpg` | | `width`, `height` | int | Image size in pixels (806 × 802) | | `date_captured` | string | Capture time, ISO 8601 | | `side` | string | `X` or `Y`, the two fiber end-faces imaged for one splice | | `pair_image_id` | int | `id` of the other side of the same splice | | `type` | string | Fiber type: `Normal` or `Trenched` | | `cores` | list | 4 cores, each `{"position_x", "position_y", "radius"}` in pixels | | `marker` | object | Orientation marker `{"position_x", "position_y"}` in pixels | **Coordinates:** the origin is the top-left corner of the image, x increases to the right and y increases downwards. The fiber centre is at about (403, 401). Example record: ```json { "id": 1, "file_name": "000001.jpg", "width": 806, "height": 802, "date_captured": "2026-03-05T18:17:22.145948", "side": "X", "pair_image_id": 2, "type": "Normal", "cores": [ {"position_x": 205, "position_y": 395, "radius": 26}, {"position_x": 405, "position_y": 199, "radius": 25}, {"position_x": 596, "position_y": 397, "radius": 23}, {"position_x": 401, "position_y": 592, "radius": 23} ], "marker": {"position_x": 668, "position_y": 256} } ``` ## Usage ### With 🤗 `datasets` ```python from datasets import load_dataset ds = load_dataset("OCEAN-Lab/MCF-Endface", split="train") sample = ds[0] sample["image"] # PIL image sample["cores"] # [{"position_x", "position_y", "radius"}, ...] sample["marker"] # {"position_x", "position_y"} ``` ### With the raw files and `MCFDataset` ```python from huggingface_hub import snapshot_download root = snapshot_download("OCEAN-Lab/MCF-Endface", repo_type="dataset", allow_patterns=["annotations.json", "images/*.jpg", "MCFDataset.py"]) import sys; sys.path.insert(0, root) from MCFDataset import MCFDataset ds = MCFDataset(root) item = ds[0] item["image"] # BGR numpy array (cv2.imread, IMREAD_UNCHANGED) item["cores"], item["marker"] # same lists/dicts as annotations.json x_item, y_item = ds.get_pair(0) # both sides of one splice ds.get_by_id(5) # look up by image id ``` `MCFDataset` needs `opencv-python`. It also needs `huggingface_hub` if you create it without a path, in which case it downloads the dataset itself. ### With plain Python ```python import json from pathlib import Path from PIL import Image data = json.load(open(Path(root) / "annotations.json")) for record in data["images"]: image = Image.open(Path(root) / "images" / record["file_name"]) cores, marker = record["cores"], record["marker"] ``` ## How the data was collected The images were captured on a fusion splicer as part of an automated MCF splicing setup. The images were taken from a Fujikura FSM 100P+ PM splicer. Each raw end-face image was cropped around the fiber cladding and resized to 806 × 802. Cores and markers were marked by an operator in a custom Python GUI before being saved. ## License [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/)