SegFly: A Dataset and 2D-3D-2D Paradigm for Aerial RGB-Thermal Semantic Segmentation at Scale

SegFly is a large-scale aerial semantic segmentation dataset featuring 20,606 high-resolution RGB images and 15,007 pixel-aligned RGB-Thermal (RGB-T) pairs. Images are captured across diverse environments and three flight altitudes (30m, 40m, 50m).

Dataset Structure

Features

Feature	Type	Description
image	Image	Raw sensor frame (RGB or LWIR Thermal)
label	Image	8-bit single-channel semantic mask mapped to 15 benchmark classes
RGB_aligned	Image	Registered RGB image (Thermal modality only; returns None for RGB modality)
scene	string	Scene identifier ("scene_01" to "scene_09")
altitude	string	Flight altitude ("30m", "40m", "50m")
modality	string	Sensor modality ("RGB" or "thermal")

Splits and Statistics

• Total Samples: 35,613 (20,606 RGB + 15,007 thermal)

Modality	Split	Scenes	Sample Count
RGB	Train	scene_01, scene_02, scene_03, scene_04, scene_05	14,738
	Val	scene_06, scene_07	1,965
	Test	scene_08, scene_09	3,842
Thermal	Train	scene_03, scene_04, scene_05	12,063
	Val/Test	scene_09	2,944

SegFly Dataset Class Mapping Reference

Class ID	Class Name	RGB Color	Color Preview
0	Unlabeled / Ignored	[0, 0, 0]	#000000
1	Road	[128, 0, 128]	#800080
2	Walkway	[204, 163, 72]	#cca348
3	Dirt	[128, 0, 0]	#800000
4	Gravel	[192, 192, 192]	#c0c0c0
6	Grass	[0, 255, 0]	#00ff00
7	Vegetation	[112, 148, 32]	#709420
8	Tree	[64, 64, 0]	#404000
9	Ground Obstacle	[255, 255, 0]	#ffff00
13	Vehicle	[0, 128, 128]	#008080
14	Water	[0, 0, 255]	#0000ff
16	Building	[255, 0, 0]	#ff0000
17	Roof	[64, 160, 120]	#40a078
33	Parking Lot	[128, 64, 128]	#804080
34	Construction	[240, 120, 120]	#f07878
36	Truck	[128, 128, 64]	#808040

How to Use

from datasets import load_dataset

# Load entire dataset
dataset = load_dataset("markus-42/SegFly")

Citation

@inproceedings{gross2026segfly,
    title={{SegFly: A Dataset and 2D-3D-2D Paradigm for Aerial RGB-Thermal Semantic Segmentation at Scale}}, 
    author={Markus Gross and Sai Bharadhwaj Matha and Rui Song and Viswanathan Muthuveerappan and Conrad Christoph and Julius Huber and Daniel Cremers},
    booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
    year={2026},
}

License

Licensed under CC BY-NC-SA 4.0.