Spatiallysaying/rwy_obb_mask2former-300-65-65

Runway Object-Oriented Bounding Box (OBB) Mask2Former Dataset v2

Dataset Description

This dataset contains aerial/satellite imagery for runway segmentation using Mask2Former architecture. The dataset supports both semantic and instance segmentation tasks, specifically designed for runway detection and delineation in aerial imagery.

Key Features

• Architecture: Optimized for Mask2Former universal segmentation
• Task Support: Semantic, instance, and panoptic segmentation
• Format: COCO-style annotations + semantic masks
• Domain: Aerial/satellite runway imagery

Dataset Statistics

Split	Images	Semantic Masks	Instance Annotations
Train	303	303	303
Val	65	65	65
Test	65	65	65

Classes

{
  "0": "_background_",
  "1": "rwy_obb"
}

Categories (COCO format)

[
  {
    "id": 0,
    "name": "_background_",
    "supercategory": "background"
  },
  {
    "id": 1,
    "name": "rwy_obb",
    "supercategory": "object"
  }
]

Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("Spatiallysaying/rwy_obb_mask2former-300-65-65")

# Load class mappings
import json
with open('id2label.json', 'r') as f:
    id2label = json.load(f)
with open('label2id.json', 'r') as f:
    label2id = json.load(f)
with open('categories.json', 'r') as f:
    categories = json.load(f)

Using with Mask2Former

from transformers import Mask2FormerImageProcessor, Mask2FormerForUniversalSegmentation
import torch
from PIL import Image
import numpy as np

# Load model and processor
processor = Mask2FormerImageProcessor.from_pretrained("facebook/mask2former-swin-large-ade-semantic")
model = Mask2FormerForUniversalSegmentation.from_pretrained("facebook/mask2former-swin-large-ade-semantic")

# Load image and semantic mask
image = Image.open("path/to/image.jpg")
semantic_mask = np.array(Image.open("path/to/semantic_mask.png"))

# Process for training
inputs = processor(images=image, segmentation_maps=semantic_mask, return_tensors="pt")

Data Format

File Structure

dataset/
├── train/
│   ├── images/           # RGB images (JPEG)
│   ├── labels/           # Semantic masks (PNG)
│   ├── annotations/      # Instance annotations (JSON)
│   └── train_annotations.json  # Complete COCO format
├── val/
│   └── ... (same structure)
├── test/
│   └── ... (same structure)
├── id2label.json        # Class ID to label mapping
├── label2id.json        # Label to class ID mapping
└── categories.json      # COCO categories format

Data Types

• Images: RGB images in JPEG format
• Semantic Labels: Grayscale segmentation masks in PNG format
  • Pixel values correspond to class IDs (0=background, 1=runway)
• Instance Annotations: COCO-format JSON files per image
  • Bounding boxes, segmentation polygons, areas, category IDs
• Resolution: Variable (original image dimensions preserved)

Annotation Format (COCO-style)

{
  "image": {
    "id": 0,
    "file_name": "image.jpg",
    "width": 1024,
    "height": 768
  },
  "annotations": [
    {
      "id": 0,
      "category_id": 1,
      "bbox": [x, y, width, height],
      "area": 12345,
      "segmentation": [[x1, y1, x2, y2, ...]],
      "iscrowd": 0
    }
  ]
}

Applications

• Semantic Segmentation: Pixel-level runway classification
• Instance Segmentation: Individual runway object detection
• Panoptic Segmentation: Combined semantic + instance understanding
• Object Detection: Runway bounding box detection
• Multi-task Learning: Universal segmentation training

Citation

@dataset{rwy_obb_mask2former_300_65_65_v2,
  title={Runway Object-Oriented Bounding Box Mask2Former Dataset v2},
  author={Spatiallysaying},
  year={2026},
  url={https://huggingface.co/datasets/Spatiallysaying/rwy_obb_mask2former-300-65-65}
}

License

MIT License