dronefreak/cabinet-mobilenetv3-small-uavid

CABiNet (MobileNetV3-Small) Finetuned on UAVid

CABiNet (MobileNetV3-Small backbone) semantic segmentation model for aerial UAV imagery using the UAVid benchmark dataset.

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

---

Performance

Metric	Score
mIoU	66.84
Pixel Accuracy	86.45
Parameters (M)	5.36
FLOPs (GFLOPs @ 1024px)	44.1

---

UAVid Model Zoo

Rank	Model	mIoU (%)	Pixel Acc (%)	Params (M)	FLOPs (GFLOPs)
1	CABiNet (MobileNetV3-Large)	68.6	87.31	9.17	54.8
2	CABiNet (MobileNetV3-Small)	66.84	86.45	5.36	44.1
3	YOLO26x-sem	64.41	85.82	40.16	430.9
4	YOLO26l-sem	63.28	84.75	17.87	192.4
5	YOLO26m-sem	61.98	84.41	14.32	152.3
6	YOLO26s-sem	61.69	84.27	6.50	44.4
7	YOLO26n-sem	58.17	82.31	1.63	11.4

---

Per-Class IoU (%)

Class	CABiNet (MobileNetV3-Large)	CABiNet (MobileNetV3-Small)	YOLO26x-sem	YOLO26l-sem	YOLO26m-sem	YOLO26s-sem	YOLO26n-sem
Clutter	69.37	67.96	67.34	65.63	64.63	63.8	61.46
Building	87.77	86.59	87.37	85.9	86.1	85.01	82.29
Road	81.62	80.94	79.82	78.87	78.61	78.14	75.25
Static Car	59.3	55.69	51.33	54.48	44.74	47.39	41.08
Tree	81.23	80.09	78.19	76.68	76.44	76.8	74.06
Vegetation	65.88	64.12	63.3	59.97	60.07	60.25	55.65
Human	30.04	27.32	21.09	19.34	20.41	18.88	15.82
Moving Car	73.62	71.99	66.81	65.38	64.88	63.3	59.73

---

Dataset

UAVid is a high-resolution UAV semantic segmentation benchmark of urban street scenes, captured from oblique aerial viewpoints along street-side flight paths.

Classes

• Clutter
• Building
• Road
• Static Car
• Tree
• Vegetation
• Human
• Moving Car

---

Usage

Install Dependencies

pip install torch huggingface_hub
git clone https://github.com/dronefreak/CABiNet.git
cd CABiNet
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -e .[dev]

Load Model from Hugging Face

from huggingface_hub import hf_hub_download
import torch

from src.models.cabinet import CABiNet

weights = hf_hub_download(
    repo_id="dronefreak/cabinet-mobilenetv3-small-uavid",
    filename="cabinet_best.pth"
)

model = CABiNet(n_classes=8, mode="small")
ckpt = torch.load(weights, map_location="cpu", weights_only=True)
state_dict = ckpt["model_state"] if isinstance(ckpt, dict) and "model_state" in ckpt else ckpt
model.load_state_dict(state_dict)
model.eval()

Run Inference

import torch

image = ...  # (1, 3, H, W) normalized float tensor
with torch.no_grad():
    logits = model(image)[0]
mask = logits.argmax(dim=1).squeeze(0).cpu().numpy()  # (H, W) class-ID map

---

Training Configuration

Setting	Value
Epochs	5000
Image size	1024
Batch size	4
Dataset	UAVid (converted images/+masks/ format)
Framework	CABiNet custom PyTorch trainer
cls_pw (class weighting)	0.5

---

Official Resources

• UAVid Semantic Segmentation Model Zoo: https://huggingface.co/collections/dronefreak/uavid-semantic-segmentation-model-zoo
• CABiNet repository: https://github.com/dronefreak/CABiNet
• CABiNet Paper: https://arxiv.org/abs/2011.00993v2
• Official UAVid Website: https://uavid.nl/
• UAVid Dataset Archive: https://doi.org/10.17026/dans-x9f-w9sa
• UAVid Paper: https://arxiv.org/abs/1810.10438
• UAVid Published Journal: https://doi.org/10.1016/j.isprsjprs.2020.05.009
• 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 UAVid dataset tooling, training/eval, and mIoU benchmarking across both. Star the repo if you find these models useful!

---

Known Limitations

Performance may degrade in:

• Very small or thin objects (e.g. pedestrians, moving cars at altitude)
• Heavy occlusion under tree canopy
• Motion blur on moving vehicles
• Mixed/very high input resolutions (UAVid source images are 3840x2160 / 4096x2160; both pipelines evaluate at reduced imgsz)

---

Citation

Please cite the following:

@article{LYU2020108,
    author = "Ye Lyu and George Vosselman and Gui-Song Xia and Alper Yilmaz and Michael Ying Yang",
    title = "UAVid: A semantic segmentation dataset for UAV imagery",
    journal = "ISPRS Journal of Photogrammetry and Remote Sensing",
    volume = "165",
    pages = "108 - 119",
    year = "2020",
    issn = "0924-2716",
    doi = "https://doi.org/10.1016/j.isprsjprs.2020.05.009",
    url = "http://www.sciencedirect.com/science/article/pii/S0924271620301295",
}

@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}
}