SegFly: Aerial RGB-Thermal Segmentation - FiftyOne subset

A grouped FiftyOne dataset — a curated subset of SegFly (Gross et al., ECCV 2026) of pixel-aligned aerial RGB-thermal (RGB-T) pairs with semantic-segmentation masks and derived instance detections.

• This dataset: Voxel51/SegFly — load directly with load_from_hub
• Loader (FiftyOne remote zoo): github.com/Burhan-Q/SegFly
• Original dataset: markus-42/SegFly · Project page · arXiv · Source code

This is an unofficial redistribution of a subset of SegFly for use with FiftyOne. All credit for the dataset belongs to the original authors (see Citation).

Installation

pip install fiftyone huggingface_hub

huggingface_hub is used to download the media from Hugging Face.

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load this dataset directly from the Hub (470 RGB-T pair groups)
dataset = load_from_hub("Voxel51/SegFly", persistent=True)

# per-class instance counts / label filtering come from the `instances` field
print(dataset.count_values("instances.detections.label"))

session = fo.launch_app(dataset)

Or load it through the FiftyOne remote zoo loader (also supports max_samples):

import fiftyone.zoo as foz

dataset = foz.load_zoo_dataset(
    "https://github.com/Burhan-Q/SegFly",
    max_samples=100,  # optional; limits the number of samples
)

Every group has a base rgb slice and a thermal slice, so the App's slice selector toggles rgb↔thermal on the same sample (like quickstart-groups). Segmentation masks render with the SegFly benchmark color scheme.

What's included (curated subset)

A ~0.6 GB set of pixel-aligned RGB-T pairs:

Scene	Altitude	Modality	Split	Groups
scene_03	30m	thermal (RGB-T pairs)	train	470

Total: 470 groups / 940 samples (470 thermal + 470 rgb slice samples). (The full SegFly release is 35,613 samples / 191 GB across 9 scenes; this dataset is the curated RGB-T pair subset only.)

Group model

One group per thermal capture, with a uniform slice set (like quickstart-groups), base slice rgb:

• slice rgb — the pixel-registered RGB frame (base)
• slice thermal — the LWIR frame

Both slices carry two label fields (sharing the aligned mask):

• ground_truth — fo.Segmentation, the semantic mask
• instances — fo.Detections derived from the mask: countable classes (Vehicle, Truck, Building, Roof, Ground Obstacle, Rock, Cable, Cable Tower, Crane, Person, Bicycle) as one detection per connected region; amorphous classes (Road, Walkway, Dirt, Gravel, Grass, Vegetation, Tree, Water, Parking Lot, Construction) as one per class. This is what enables App per-class filtering and per-class instance counts (the semantic Segmentation field alone cannot be filtered/counted by class).

Because every group has both slices, toggling rgb↔thermal in the App stays on the same sample. Per-sample fields: scene, altitude, modality. Split is a sample tag. Note: instances are derived from the semantic masks via connected components (not source instance annotations); "stuff" classes are stored as a single region per image.

Reusing the instance derivation (e.g. on the full SegFly release)

The instances field ships precomputed in this dataset. The derivation is also exposed as reusable functions in the loader repo, so you can apply the same stuff/thing logic to any SegFly semantic mask (including the full 191 GB markus-42/SegFly):

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

import segfly  # from github.com/Burhan-Q/SegFly (add the cloned repo dir to your path)

# a single mask -> instance detections
dets = segfly.segmentation_to_instances(sample["ground_truth"])

# or populate an `instances` field across a whole dataset (grouped or flat)
full = load_from_hub(
    "markus-42/SegFly",
    format="ParquetFilesDataset",
    ...,
)
segfly.add_instances(full)  # in_field="ground_truth", out_field="instances"

print(full.count_values("instances.detections.label"))

MASK_TARGETS (class map) and MASK_TYPES (the stuff/thing split) are module-level constants you can inspect or override.

Classes

The stored masks contain the raw OccuFly class IDs (0–36). mask_targets names every ID that can appear:

ID	Name		ID	Name		ID	Name
0	Unlabeled		8	Tree		17	Roof
1	Road		9	Ground Obstacle		21	Cable
2	Walkway		10	unknown_10		22	Cable Tower
3	Dirt		11	Person		33	Parking Lot
4	Gravel		12	Bicycle		34	Construction
5	Rock		13	Vehicle		35	Crane
6	Grass		14	Water		36	Truck
7	Vegetation		16	Building

Note. SegFly's published "15 benchmark classes" are a documented post-processing remap that is not baked into the mask files: Rock(5) and Cable Tower(22) → Ground Obstacle(9); Person(11), Bicycle(12), Cable(21), Crane(35) → Unlabeled(0). Apply this remap if you need the benchmark protocol. ID 10 appears in the data but is undocumented in the source and is left un-named as unknown_10.

License & attribution

SegFly is released under CC BY-NC-SA 4.0 (non-commercial, share-alike, attribution). This redistribution keeps the same license. Use is non-commercial only; you must attribute the original authors and share derivatives under the same terms.

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},
}