CFC26 Dataset Card

Dataset Summary

CFC26 is a large-scale benchmark dataset for fish detection, tracking, and counting in underwater ARIS sonar video. It is designed to evaluate generalization under distribution shift and deployment-relevant performance, particularly in ecologically diverse and acoustically challenging environments.

The dataset spans multiple river systems with substantial variation in fish size, density, sonar range, and background structure, and introduces evaluation settings that reflect real deployment conditions. It also supports analysis of directional counting accuracy, which is critical for estimating net fish passage in ecological monitoring.

---

Dataset Details

Property	Value
Task	Fish detection, multi-object tracking, counting
Sensor	ARIS Explorer
Annotation format	COCO JSON, MOT
Number of locations	11 deployment locations across 9 river systems
Clips	4,363
Bounding boxes	1.0M+
Tracks	14,072
License	cc-by-4.0

Splits are assigned temporally per location using natural gaps in recording to form chronologically ordered, non-overlapping train, validation, and test sets. This prevents near-duplicate temporal contexts from appearing across splits and provides a more realistic estimate of deployment performance, where models are applied to future unseen data rather than randomly sampled frames.

---

Motivation

Monitoring fish passage at rivers is critical for fisheries management and conservation. Manual review of sonar video is labor-intensive and does not scale to the volume of data collected during migration seasons. CFC26 provides large-scale annotated data to support automated fish detection and counting systems that can operate directly on raw ARIS sonar imagery.

---

Locations

Data were collected at 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

Kenai locations are further stratified by stratum and bank position (channel, left bank, right bank), reflecting distinct sonar deployment configurations at that site.

---

Dataset Structure

File Layout

{location}/
└── {clip_name}/
    └── *.png / *.jpg       # Each frame is a derived 3-channel representation constructed from raw sonar frames

coco/
└── {location}/
    └── {clip_name}.json    # COCO annotation file per clip

mot/
└── {location}/
    └── {clip_name}    
      └── gt.txt            # MOT annotation file per clip

Clip Naming Convention

Image filenames encode clip identity and frame index:

{source}_{YYYY-MM-DD}_{HHMMSS}_{start_frame}_{end_frame}_{frame_index}.jpg

Timestamps in filenames are used for temporal train/val/test assignment.

---

Annotation Format

Annotations are provided in COCO JSON format (primary) and MOT text format.

COCO JSON

Each clip has a corresponding JSON file with the standard COCO structure:

Image fields:

Field	Description
id	Unique image ID
file_name	Frame filename
width, height	Frame dimensions in pixels

Annotation fields:

Field	Description
bbox	[x, y, width, height] in pixels (top-left origin)
track_id	(when present) identifies object ID consistency within a clip
category_id	Fish category ID
image_id	Corresponding frame ID

Categories:

Name	Description
FISH	Confident fish detection

MOT Format

One row per detection per frame:

frame_id, track_id, left, top, width, height, conf, x, y, z

Note: frame_id is 1-based and conf is set to 1 for all annotations.

Train / Val / Test Splits

Splits are assigned temporally per location: clips are ordered chronologically and divided into contiguous train, validation, and test windows. Chronological ordering within each location is derived from timestamps encoded in clip filenames. This design evaluates generalization across time within a deployment season — a realistic condition for monitoring applications — rather than random frame-level splits that would inflate performance estimates.

Split files are provided per location as text files (train.txt, val.txt, test.txt) under splits/{location}/. Each line corresponds to a COCO annotation filename (including the .json extension), which identifies a clip.

---

Example Usage

Below is a minimal example of loading frames and COCO-style annotations for a single clip.

import json
from pathlib import Path
from PIL import Image

# paths
data_root = Path("CFC26")
location = "Klamath"
clip = "example_clip"

# load annotations
ann_path = data_root / "labels" / "coco" / location / f"{clip}.json"
with open(ann_path) as f:
    coco = json.load(f)

# map image_id to file
# clip refers to the clip directory name (without file extension)
images = {img["id"]: img for img in coco["images"]}

# clip image directory
img_dir = data_root / location / clip 

for ann in coco["annotations"]:
    img_info = images[ann["image_id"]]
    img_path = img_dir / img_info["file_name"]

    if not img_path.exists():
        stem = Path(img_info["file_name"]).stem
        candidates = list(img_dir.glob(f"{stem}.*"))

        if len(candidates) == 0:
            raise FileNotFoundError(f"No image found for: {stem}")

        img_path = candidates[0]

    image = Image.open(img_path)

    bbox = ann["bbox"]  # [x, y, w, h]
    track_id = ann.get("track_id", None)
    category_id = ann["category_id"]

    # --- your processing here ---

Data Collection

Sonar units were deployed at each site and recorded continuously or during active monitoring windows. Clips were segmented from raw recordings based on temporal structure.

Annotations were produced using a structured pipeline designed to ensure completeness and consistency. Annotators were provided with both raw video and echogram visualizations, which collapse temporal sonar data into 2D representations where fish appear as trajectories. This aids detection of small or low-contrast fish without altering annotation definitions.

---

Limitations

• Class imbalance. Empty frames (no fish present) constitute a large fraction of the data at most locations. Models must handle significant negative-to-positive imbalance.
• Domain shift across locations. Sonar imagery varies considerably across deployment sites due to differences in water turbidity, fish size, fish density, flow rate, camera angle, and hardware configuration. Models trained on a subset of locations may not generalize without adaptation.
• Intrinsic difficulty variation. Some locations remain challenging even under in-distribution training due to sensing conditions such as fish size and sonar range. This reflects real deployment constraints and may limit achievable performance independent of model design.

---

Ethical Considerations

CFC26 contains no personally identifiable information. Data were collected in collaboration with fisheries agencies and organizations. Use of this dataset for automated fish monitoring should be accompanied by appropriate ecological expertise and should not be used as a sole basis for management decisions without validation by domain experts.

---

Citation

(Citation will be updated with full author list and proceedings details upon publication.)

---

License

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

---

Mini Sample

A curated mini version of CFC26 (CFC26-Mini) is available for demonstration purposes. It contains 1 clip per split per location with 10 frames per clip (only positive samples).