CFC26-Mini Dataset Card

Dataset Summary

CFC26-Mini is a curated sample of the full CFC26 dataset — a large-scale benchmark for fish detection and counting in underwater ARIS sonar video collected across 9 field locations in North America. This mini version is intended for demonstration purposes.

Note: This is a sample of the full CFC26 dataset. The full dataset contains thousands of clips and hundreds of thousands of annotated frames across all locations and splits. See Full Dataset for details.

---

Dataset Details

Property	Value
Task	Fish detection, counting
Annotation format	COCO JSON (bounding boxes + track IDs)
License	CC BY 4.0

---

Sampling Methodology

CFC26-Mini is constructed from the full CFC26 dataset using the following procedure:

• 1 clip sampled per split (train / val / test) per location
• 10 frames sampled per clip
• Only frames containing at least one fish annotation are included (positive samples only)
• Temporal train/val/test splits are inherited from the full CFC26 dataset splits

---

Dataset Structure

cfc26-mini/
├── {location}/
│   └── {clip_name}/
│       └── *.png / *.jpg          # 3-channel representations (fish-positive only)

Each location directory contains up to 3 clip subdirectories (one per split). Frame filenames encode the clip identity and frame index:

{clip_name}_{frame_index}.png  # png and/or jpg

---

Locations

The mini dataset includes samples from the following field sites:

Location	Region
Chemainus	British Columbia, Canada
Dungeness	Washington, USA
Eel	California, USA
Elwha	Washington, USA
Channel	Alaska, USA
Kenai	Alaska, USA
Rightbank	Alaska, USA
Klamath	California, USA
Mad	California, USA
Nanaimo	British Columbia, Canada
Nushagak	Alaska, USA

---

Data Fields

Each frame is a 3-channel representation made up of the raw frame, the background-subtracted frame, and the difference between consecutive background-subtracted frames. Annotations are provided in COCO JSON format alongside the full dataset.

COCO annotation fields per bounding box:

Field	Description
bbox	[x, y, width, height] in pixels (top-left origin)
track_id	Unique fish identity within a clip
category_id	Fish category (FISH)
image_id	Links to the corresponding frame

---

Full Dataset

The full CFC26 dataset contains:

• Locations across Alaska, California, Washington, and British Columbia
• Thousands of annotated clips spanning multi-week deployments per site
• Temporal train/val/test splits designed to evaluate generalization across time
• Complete COCO JSON annotations with bounding boxes and track IDs
• Per-clip ARIS sonar metadata (frame rate, pixel scale, window parameters)

The full dataset is available at: (link to be added upon release)

---

Limitations

• Positive samples only. CFC26-Mini contains exclusively fish-positive frames (frames with at least one fish annotation). Empty frames — which constitute a significant portion of the full dataset — are not included. Models trained or evaluated on this mini dataset will not be exposed to the true class imbalance present in real deployments.
• 1 clip per split per location. A single clip is unlikely to be representative of the full temporal and environmental variation within a location.
• 10 frames per clip. Frames are randomly sampled for this demonstration and are not sequential. Temporal context (motion, track continuity) is not preserved.
• Not for benchmarking. Results on CFC26-Mini are not comparable to results reported on the full CFC26 dataset. Use the full dataset for all benchmark evaluation.

---

Citation

(Citation will be updated with full details upon publication.)

---

License

CFC26-mini is released under the Creative Commons Attribution 4.0 International CC BY 4.0 License.