TopAir: Light-Weight Aerial dataset for Monocular Depth Estimation and Semantic Segmentation

• Curated by: Yara AlaaEldin
• Shared by: MaLga Lab

This dataset was collected with a nadir view (top view) using AirSim simulator in various outdoor environments (both rural and urban). It contains video frames collected in differet trajectories along with annotation of depth maps and semantic segmentation maps. The extrinsic camera locations corresponding to each frame are also provided. TopAir was collected at low, mid, and high altitudes varying from 10 to 100 meters above the ground. This is a light-weight dataset with an image resolution of 384x384 and 10,385 total number of frames.

Dataset Details

The dataset was collected using different environments in UnrealEngine4 integrated with AirSim simulator. These environments are:

• Africa (3 trajectories)
• City Park (7 trajectories)
• Oak Forest (3 trajectories)
• Rural Australia (4 trajectories)
• Assetsville Town (15 trajectories)
• Accustic Cities (3 trajectories)
• Downtown City (7 trajectories)
• Edith Finch (2 trajectories)
• Landscape Mountains (1 trajectory)
• Nature Pack (1 trajectory)
• Neighbourhood (8 trajectories)

The semantic segmentation maps of TopAir contain 9 classes:

Class ID	RGB Color Palette	Class Name	Definition
0	(127, 175, 230)	Sky	the sky
1	(75, 163, 185)	Water	Includes all water surfaces
2	(50, 128, 0)	Trees	all trees and vegetation above the ground
3	(232, 250, 80)	Land	all ground types (dirt ground, rocky ground, ground vegetation, sand, etc)
4	(237, 125, 49)	Vehicle	all types of ground vehicles (car, trucks, ..)
5	(70, 70, 70)	Rocks	boulders and rocks
6	(142, 1, 246)	Road	Lanes, streets, paved areas on which cars drive
7	(255, 128, 128)	Building	buildings, residential houses, and constructions
8	(128, 64, 64)	Others	any other objects not incuded in the above classes

Per-pixel class distribution:

To convert depth maps to their corresponding meter values:

depth_value (in meters) = pixel_value*100.0/255.0

The reference frames convention used in the collection of TopAir:

Dataset Sources

• Paper: ....
• Demo: ....

Uses

This dataset can be used in monocular depth estimation and semantic segmentation tasks concrened with aerial applications in outdoor environments.

Dataset Structure

The dataset is organized as follows:

├── AccuCities_1 (environmentName_trajectoryNum)
│   ├── depth (depth maps annotation)
│   ├── images (RGB frames)
│   ├── seg_colored (segmentation maps in colors)
│   ├── seg_id (segmentation maps represented with class ids)
│   └── camera_loc.txt (camera locations and orientations)
│   
├── AccuCities_2
│   ├── depth
│   ├── images
│   ├── seg_colored
│   ├── seg_id
│   └── camera_loc.txt
│
├── AccuCities_2
│   ├── depth
│   ├── images
│   ├── seg_colored
│   ├── seg_id
│   └── camera_loc.txt
│
├── ....
│
└── RuralAust3_2
    ├── depth
    ├── images
    ├── seg_colored
    ├── seg_id
    └── camera_loc.txt

Data Collection and Processing

......

Citation

If you use TopAir dataset, please like our dataset repository and cite this page:

BibTeX:

@misc{huggingfaceYaraalaa0TopAirDatasets,
    author = {},
    title = {yaraalaa0/{T}op{A}ir · {D}atasets at {H}ugging {F}ace --- huggingface.co},
    howpublished = {\url{https://huggingface.co./datasets/yaraalaa0/TopAir}},
    year = {},
    note = {[Accessed 27-02-2025]},
}

APA:

AlaaEldin, Y. (n.d.). Yaraalaa0/TopAir · Datasets at hugging face. yaraalaa0/TopAir · Datasets at Hugging Face. https://huggingface.co./datasets/yaraalaa0/TopAir