README.md

metadata

license: apache-2.0
task_categories:
  - object-detection
tags:
  - multi-domain

Multi-Domain Perception Dataset

Unlabeled raw image dataset organized across domain environments, perspectives, and source origins. Designed for downstream annotation (e.g., Label Studio) and computer vision pipelines.

Repository Structure

Data files are organized cleanly by domain, provenance type, and collection date to prevent naming collisions and facilitate subset loading:

data/
├── <domain>/
│   └── <source_type>/
│       └── <date_collected>/
│           └── <raw_data_files>
└── metadata.jsonl

Dataset Schema (metadata.jsonl)

Each line in metadata.jsonl tracks individual file paths alongside collection provenance:

{
  "file_name": "data/underwater/real_world/2026-08-19/camera_front/frame_0001.webp",
  "metadata": {
    "author": "Collector Name",
    "domain": "Underwater",
    "perspective": "Planar",
    "date_collected": "2026-08-19",
    "notes": "Field trial under ambient light",
    "md5": "e99a18c428cb38d5f260853678922e03",
    "provenance": {
      "type": "real_world",
      "platform": "sensor_rig"
    }
  }
}

Supported Fields

• Domain: Surface, Ground, Underwater, Aerial, Other
• Perspective: Horizon (Looking to fore/aft/sides), Nadir (Looking down), Zenith (Looking up), Not Applicable
• Source Type: real_world, synthetic, simulated

Usage with Hugging Face datasets

from datasets import load_dataset

# Load full dataset with metadata features
ds = load_dataset("username/repo-name")

# Filter for real-world underwater data
underwater_real = ds.filter(
    lambda x: x["metadata"]["domain"] == "Underwater" 
    and x["metadata"]["provenance"]["type"] == "real_world"
)