dronefreak/aeroscapes-yolo26x-sem

YOLO26x-sem Finetuned on AeroScapes

Fine-tuned YOLO26x semantic segmentation model for aerial drone imagery using the AeroScapes benchmark dataset.

This model is part of the AeroScapes Semantic Segmentation Model Zoo, a collection of CABiNet and YOLO26 models trained and evaluated under a common pipeline for aerial semantic segmentation.

_{Qualitative results on AeroScapes val-split examples — single-scale (imgsz=896) inference, no TTA.}

---

Performance

Metric	Score
mIoU	68.36
Pixel Accuracy	93.91
Parameters (M)	40.15
FLOPs (GFLOPs @ 720px)	213.0

---

AeroScapes Model Zoo

Rank	Model	mIoU (%)	Pixel Acc (%)	Params (M)	FLOPs (GFLOPs)
1	YOLO26x-sem	68.36	93.91	40.15	213.0
2	YOLO26l-sem	68.0	93.16	17.86	95.1
3	CABiNet (MobileNetV3-Large)	67.83	94.32	9.18	27.4
4	YOLO26m-sem	66.97	93.51	14.31	75.2
5	YOLO26s-sem	65.44	93.22	6.50	21.9
6	YOLO26n-sem	64.86	92.84	1.63	5.6

---

Per-Class IoU (%)

Class	YOLO26x-sem	YOLO26l-sem	CABiNet (MobileNetV3-Large)	YOLO26m-sem	YOLO26s-sem	YOLO26n-sem
Background	81.69	79.43	82.67	80.75	79.31	78.43
Person	53.34	49.28	46.9	50.81	49.44	44.51
Bike	35.61	42.18	42.35	38.4	30.09	35.7
Car	86.16	87.67	85.08	84.76	84.54	85.28
Drone	60.46	65.97	61.77	58.91	59.73	50.88
Boat	82.04	79.76	77.67	80.33	76.19	80.28
Animal	41.21	37.08	33.74	38.22	36.21	37.15
Obstacle	26.91	29.17	23.13	24.14	19.8	17.29
Construction	75.76	67.09	78.81	72.87	71.02	74.39
Vegetation	94.07	93.6	94.3	93.99	93.93	93.55
Road	92.14	92.07	92.91	91.94	91.64	89.1
Sky	90.96	92.72	94.57	88.57	93.35	91.73

---

Evaluation Visualizations

Per-Class IoU Bar Chart

Confusion Matrix

Loss Curves

---

Dataset

AeroScapes is an aerial semantic segmentation benchmark of ~3,200 720p images captured by a fleet of drones across roughly 10 flight sequences, covering urban and suburban scenes with people, vehicles, and small objects.

Classes

• Background
• Person
• Bike
• Car
• Drone
• Boat
• Animal
• Obstacle
• Construction
• Vegetation
• Road
• Sky

---

Usage

Install Dependencies

pip install ultralytics huggingface_hub

Load Model from Hugging Face

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

weights = hf_hub_download(
    repo_id="dronefreak/aeroscapes-yolo26x-sem",
    filename="best.pt"
)

model = YOLO(weights)

Run Inference

results = model.predict(source="image.png", task="semantic", imgsz=896)
mask = results[0].semantic_mask.cpu().numpy().data  # (H, W) class-ID map

---

Training Configuration

Setting	Value
Epochs	120
Image size	896
Batch size	8
Dataset	AeroScapes (converted images/+masks/ format)
Framework	Ultralytics YOLO
cls_pw (class weighting)	0.5

---

Official Resources

• AeroScapes Semantic Segmentation Model Zoo: https://huggingface.co/collections/dronefreak/aeroscapes-semantic-segmentation-model-zoo
• CABiNet repository: https://github.com/dronefreak/CABiNet
• CABiNet Paper: https://arxiv.org/abs/2011.00993v2
• AeroScapes Dataset (Hugging Face): https://huggingface.co/datasets/dronefreak/Aeroscapes
• AeroScapes Repository: https://github.com/ishann/aeroscapes
• AeroScapes Paper (CMU PDF): https://www.cs.cmu.edu/~deva/papers/aeroscapes.pdf
• Ultralytics YOLO: https://github.com/ultralytics/ultralytics
• Ultralytics YOLO26 Paper: https://arxiv.org/abs/2606.03748

---

Training Framework

Trained with the CABiNet repository, which pairs its own real-time segmentation trainer with a parallel Ultralytics YOLO26-sem pipeline — shared dataset tooling, training/eval, and mIoU benchmarking across UAVid, AeroScapes, and VDD. Star the repo if you find these models useful!

---

Known Limitations

Performance may degrade in:

• Very rare classes (Bike, Drone, Boat) each cover well under 2% of any single frame in the val split — expect unreliable predictions for these
• No official test split — all reported numbers are on the val split, unlike UAVid/VDD
• Frames are drawn from only ~10 flight sequences, so scene diversity is lower than UAVid/VDD despite the image count
• Native resolution (1280x720) downsampled to the eval imgsz — fine detail on small objects (pedestrians, bikes) can be lost

---

Citation

Please cite the following:

@inproceedings{nigam2018ensemble,
  title={Ensemble Knowledge Transfer for Semantic Segmentation},
  author={Nigam, Ishan and Huang, Chen and Ramanan, Deva},
  booktitle={2018 IEEE Winter Conference on Applications of Computer Vision (WACV)},
  year={2018}
}

@INPROCEEDINGS{9560977,
  author={Kumaar, Saumya and Lyu, Ye and Nex, Francesco and Yang, Michael Ying},
  booktitle={2021 IEEE International Conference on Robotics and Automation (ICRA)},
  title={CABiNet: Efficient Context Aggregation Network for Low-Latency Semantic Segmentation},
  year={2021},
  pages={13517-13524},
  doi={10.1109/ICRA48506.2021.9560977}
}

@article{Kumaar_Real-time_Semantic_Segmentation_2021,
  author = {Kumaar, Saumya and Lyu, Ye and Nex, Francesco and Yang, Michael Ying},
  doi = {10.1016/j.isprsjprs.2021.06.006},
  journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
  pages = {124--134},
  title = {{Real-time Semantic Segmentation with Context Aggregation Network}},
  url = {https://www.sciencedirect.com/science/article/pii/S0924271621001647},
  volume = {178},
  year = {2021}
}

@article{jocher2026ultralytics,
  title={Ultralytics YOLO26: Unified Real-Time End-to-End Vision Models},
  author={Jocher, Glenn and Qiu, Jing and Liu, Mengyu and Lyu, Shuai and Akyon, Fatih Cagatay and Kalfaoglu, Muhammet Esat},
  journal={arXiv preprint arXiv:2606.03748},
  year={2026}
}

@software{cabinet_uavid_benchmark,
  author = {Kumaar, Saumya},
  title = {CABiNet: Semantic Segmentation Benchmarking on UAVid (CABiNet vs. YOLO26)},
  url = {https://github.com/dronefreak/CABiNet},
  year = {2026}
}