xjh19972/Robotic-Fish-Pose-Dataset

Robotic Fish Pose Dataset

Dataset Description

A pose estimation dataset for robotic fish, annotated in COCO format with bounding boxes and 3 keypoints per fish instance: head, body, and tail.

Dataset Structure

Splits

Split	Images	Annotations	Source
train_images	818	3,521	Labeled still images
train_videos	301	1,505	Labeled video frames
test	200	1,000	Test video frames

Additionally, 2 unlabeled .avi video files are included in data/unlabeled_videos/ for semi-supervised learning.

Annotation Format

Annotations follow the COCO keypoint detection format. Each image has a metadata.jsonl entry with an annotations field containing a JSON list of objects:

• Bounding box: [x, y, width, height] in pixels
• Keypoints: [x1, y1, v1, x2, y2, v2, x3, y3, v3] where each keypoint has (x, y, visibility)
  • Keypoint 1: Head
  • Keypoint 2: Body
  • Keypoint 3: Tail
  • Visibility: 0 = not labeled, 1 = labeled but not visible, 2 = labeled and visible
• Skeleton: [[1, 2], [2, 3]] (head-body, body-tail)

Data Fields

• image (Image): The image file
• image_id (string): Unique identifier for the image
• width (int): Image width in pixels (752)
• height (int): Image height in pixels (480)
• annotations (string): JSON-encoded list of fish annotations, each containing:
  • id (int): Annotation ID
  • bbox (list[float]): Bounding box [x, y, w, h]
  • area (float): Bounding box area
  • keypoints (list[float]): 9 values for 3 keypoints
  • num_keypoints (int): Number of labeled keypoints
  • iscrowd (int): 0 for individual instances

Usage

from datasets import load_dataset
import json

# Load from HuggingFace Hub
dataset = load_dataset("xjh19972/Robotic-Fish-Pose-Dataset")

# Access splits
train_images = dataset["train_images"]
train_videos = dataset["train_videos"]
test = dataset["test"]

# Inspect a sample
sample = train_images[0]
print(sample["image_id"])
annotations = json.loads(sample["annotations"])
for ann in annotations:
    print(f"  bbox: {ann['bbox']}, keypoints: {ann['keypoints']}")

Directory Structure

Robotic-Fish-Pose-Dataset/
├── README.md
├── .gitattributes
├── data/
│   ├── train_images/               # 818 labeled still images (.jpg) + metadata.jsonl
│   ├── train_videos/               # 301 labeled video frames (.jpg) + metadata.jsonl
│   ├── test/                       # 200 test video frames (.jpg) + metadata.jsonl
│   └── unlabeled_videos/           # 2 unlabeled .avi files