README.md

metadata

pretty_name: 'FishCut: Industrial Headless-Hake Cut-Line Localization'
license: cc-by-nc-4.0
language:
  - en
task_categories:
  - keypoint-detection
  - object-detection
  - image-feature-extraction
annotations_creators:
  - expert-generated
size_categories:
  - n<1K
tags:
  - food
  - fish
  - hake
  - seafood
  - cut-line
  - cutting-point
  - fish-processing
  - industrial
  - computer-vision
  - keypoint-detection
  - industry-4.0
  - smart-manufacturing
  - food-informatics
  - sustainability
dataset_info:
  features:
    - name: image
      dtype: image
    - name: image_id
      dtype: string
    - name: bbox
      list: float32
    - name: cut_line
      list:
        list: float32
  splits:
    - name: train
      num_bytes: 41004650
      num_examples: 370
    - name: validation
      num_bytes: 27855903
      num_examples: 247
  download_size: 68866893
  dataset_size: 68860553
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*

🐟 The FishCut Dataset: Cut-Line Localization for Industrial Headless-Hake Steaking

This dataset supports vision-based cut-line localization for industrial fish steaking. In a headless-hake processing line, an operator must place a reference cut on the head side of each fish: a cut placed too close to the head-side boundary risks including bone fragments, while a cut placed too far into the body wastes high-value muscle. FishCut was collected to automate this expert decision with deep learning.

The dataset consists of 617 expert-annotated images of headless hake trunks, captured on an active seafood production line with an Intel RealSense D456 camera. This release provides the cropped fish-ROI version, in which each image is cropped around the fish trunk. The dataset is intended for keypoint detection and coordinate-regression approaches to cut-line localization.

🗂️ automated cut-line localization (red = Prediction , green= Ground truth)

Each instance contains:

• image: an RGB image of a headless hake trunk (cropped fish ROI).
• keypoints: the reference cut-line, given as four values [x_b, y_b, x_t, y_t] — the bottom and top interface points of the cut-line, normalized to [0, 1] by the image width and height.
• image_id: a unique identifier for each sample.

🏷️ Annotation Format

The reference cut-line is defined by two keypoints (the bottom and top of the cut), stored as four normalized coordinates [x_b, y_b, x_t, y_t]. The ground-truth cut-lines were defined by three expert operators with extensive experience in manual hake steaking, who labelled each image jointly by consensus, producing a single mutually agreed annotation per image.

Note: coordinates are normalized to the dimensions of each (cropped) image, so multiply by the image width/height to recover pixel positions.

🧪 How to read and display examples

from datasets import load_dataset
from PIL import Image, ImageDraw

# --- Configuration ---
DATASET_NAME = "salahkhenfer/FishCut"
SAMPLE_INDEX = 0
OUTPUT_IMAGE = "annotated_FishCut.png"

if __name__ == "__main__":
    dataset = load_dataset(DATASET_NAME)
    split = "train" if "train" in dataset else list(dataset.keys())[0]
    sample = dataset[split][SAMPLE_INDEX]

    image = sample["image"].convert("RGB")
    width, height = image.size
    draw = ImageDraw.Draw(image)

    # keypoints = [x_b, y_b, x_t, y_t], normalized to [0, 1]
    xb, yb, xt, yt = sample["keypoints"]
    p_bottom = (xb * width, yb * height)
    p_top = (xt * width, yt * height)

    # draw the reference cut-line in green
    draw.line([p_bottom, p_top], fill=(0, 255, 0), width=3)

    image.save(OUTPUT_IMAGE)
    print(f"Annotated image saved as {OUTPUT_IMAGE}")

🤝 Partners

📖 Citation

@article{khenfer2026FishCut,
  title   = {Artificial Intelligence and Computer Vision for Automated Cut-Line Localization in Industrial Fish Processing},
  author  = {Salah E.Khenfer,  Mohamed L.Mekhalfi, Mingdeng Shi,Hua Zou},
  journal = {Future Foods},
  year    = {2026}
}