hojayfa-sir/European_Green_Crab

Multi-Species Crab Detection (Combined 5ds)

An aggregated and augmented marine object detection dataset designed for multi-class identification of three crab species: European Green Crab, Rock Crab, and Jonah Crab.

The dataset was constructed by merging five independent, open crab datasets into a single unified COCO-formatted collection, followed by balanced train/val/test splitting, bounding-box standardization, and a 5× training augmentation pipeline.

---

Dataset Details

Source Dataset Origins

The core data was compiled by merging 5 open-source source datasets:

Source Dataset Name	Base Image Count	Verified Public Source
European crab	2,364	Roboflow Universe: european-crab-wepdb
Crab_Train	269	Private / Archived Universe Source
crab-detection	126	Roboflow Universe: crab-detection-iqurg
Crab	82	Private / Archived Universe Source
Crab Rov	48	Private / Archived Universe Source
Total Base Pool	2,889	—

---

Dataset Structure & Statistics

Base Class Distribution (Pre-Augmentation)

The raw unaugmented pool of 2,889 images contains 6,630 annotated crab instances across the three target classes:

Class Name	Train Annotations	Validation Annotations	Test Annotations	Total Annotations
European Green Crab	1,780	500	277	2,557
Rock Crab	1,538	417	249	2,204
Jonah Crab	1,285	358	226	1,869
Total	4,603	1,275	752	6,630

Annotation Density & Image Dimensions

• Annotation Density: An average of 2.3 annotations per image.
  • Single-object scenes: 2,405 images
  • 2 to 5 objects: 120 images
  • 6 to 9 objects: 133 images
  • 10+ dense clusters: 231 images
• Raw Spatial Dimensions:
  • Median image resolution: 386 × 342 pixels
  • Average image size: ~0.13 MP (ranging from 0.09 MP to 12.19 MP)
  • Aspect ratio distribution: 87.0% wide, 9.2% very wide, 3.9% tall

---

Preprocessing & Augmentation Pipeline

Augmentation was applied exclusively to the training split to introduce environmental variations (turbid water, motion artifacts, orientation changes) while leaving validation and test splits uncorrupted.

Preprocessing

• Auto-Orient: EXIF orientation flags stripped and applied.
• Standardized Letterboxing: Images scaled and padded with white borders to fit a uniform 576 × 576 canvas, preventing bounding box aspect-ratio distortion.

Training Augmentations (5× Expansion)

• Rotation: Uniform random rotation between -15° and +15°
• Brightness & Exposure: Random photometric shifts between -15% and +15%
• Sensor Noise: Salt-and-pepper noise injected up to 2% of pixels
• Motion Blur: 20px blur kernel at 45° across 2 simulated frames

Final Split Breakdown

Split	Images	Split Percentage	Source Notes
Train	10,110	92.1%	Augmented 5× from base training images
Validation	578	5.3%	Standardized raw letterboxed images
Test	289	2.6%	Standardized raw letterboxed images
Total	10,977	100%	Full dataset volume

---

Directory & File Layout

When downloaded or cloned from Hugging Face, the dataset follows the standard COCO structure:

x5-ds-combined/
├── train/
│   ├── _annotations.coco.json
│   └── *.jpg
├── valid/
│   ├── _annotations.coco.json
│   └── *.jpg
└── test/
    ├── _annotations.coco.json
    └── *.jpg

Annotation Schema (_annotations.coco.json)

Each annotation file adheres strictly to standard COCO bounding box conventions:

{
  "images": [
    {
      "id": 0,
      "file_name": "example_001.jpg",
      "width": 576,
      "height": 576
    }
  ],
  "annotations": [
    {
      "id": 0,
      "image_id": 0,
      "category_id": 1,
      "bbox": [x_min, y_min, width, height],
      "area": 14250.0,
      "iscrowd": 0
    }
  ],
  "categories": [
    {"id": 0, "name": "crabs", "supercategory": "none"},
    {"id": 1, "name": "European Green Crab", "supercategory": "crabs"},
    {"id": 2, "name": "Rock Crab", "supercategory": "crabs"},
    {"id": 3, "name": "Jonah Crab", "supercategory": "crabs"}
  ]
}

Note on Category IDs: Depending on the tool used, category index 0 may represent a generic supercategory placeholder. When loading with libraries like supervision or pycocotools, ensure your class mapper aligns with IDs 1..3.

---

Loading the Dataset in Python

Using supervision

import supervision as sv

dataset = sv.DetectionDataset.from_coco(
    images_directory_path="test/",
    annotations_path="test/_annotations.coco.json"
)

print(f"Loaded {len(dataset)} test images across classes: {dataset.classes}")

Using Hugging Face datasets

from datasets import load_dataset

# Loaded directly from Hugging Face repository
ds = load_dataset("hojayfa-sir/European_Green_Crab")
print(ds)

---

Intended Uses & Limitations

Suitable Use Cases

• Marine wildlife monitoring and underwater ROV autonomous survey pipelines.
• Invasive species detection (specifically tracking the spread of the invasive European Green Crab).
• Benchmarking transformer detectors (such as RF-DETR, RT-DETR, and Co-DETR) on small-to-medium biological targets.

Limitations

• Visual similarities between juvenile Rock Crabs and Jonah Crabs can result in slight label noise under poor illumination conditions.
• Images originate from mixed lighting environments (direct lab photography, wet market tables, and submerged underwater ROV video frames). Additional domain adaptation may be needed for specific deployment depths.

---

Citation & Attribution

If you use this aggregated dataset in your research or application, please cite the underlying open Roboflow sources. The actual dataset release can be found on Roboflow here.

@misc{crab_combined_5ds,
  title={Multi-Species Crab Detection (Combined 5ds)},
  author={Abu Hojayfa},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/hojayfa-sir/European_Green_Crab}}
}