dronefreak/vdd-cabinet-mobilenetv3-large

CABiNet (MobileNetV3-Large) Finetuned on VDD

CABiNet (MobileNetV3-Large backbone) semantic segmentation model for aerial drone imagery using the VDD (Varied Drone Dataset) benchmark dataset.

This model is part of the VDD 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 VDD test-split examples — single-scale (imgsz=1024) inference, no TTA.}

---

Performance

Metric	Score
mIoU	77.76
Pixel Accuracy	89.57
Parameters (M)	9.17
FLOPs (GFLOPs @ 1024px)	54.8

---

VDD Model Zoo

Rank	Model	mIoU (%)	Pixel Acc (%)	Params (M)	FLOPs (GFLOPs)
1	YOLO26x-sem	78.83	89.79	40.16	430.9
2	YOLO26l-sem	78.57	89.68	17.87	192.4
3	CABiNet (MobileNetV3-Large)	77.76	89.57	9.17	54.8
4	YOLO26m-sem	77.02	88.3	14.32	152.3
5	YOLO26s-sem	76.35	88.27	6.50	44.4
6	YOLO26n-sem	73.99	86.32	1.63	11.4

---

Per-Class IoU (%)

Class	YOLO26x-sem	YOLO26l-sem	CABiNet (MobileNetV3-Large)	YOLO26m-sem	YOLO26s-sem	YOLO26n-sem
Other	64.15	65.88	66.28	61.04	60.99	57.27
Wall	69.26	70.93	65.87	70.06	67.48	64.76
Road	72.61	72.18	70.2	70.78	69.68	67.78
Vegetation	90.14	89.68	91.06	89.59	88.63	85.01
Vehicle	70.99	68.52	73.38	68.76	66.08	62.96
Roof	89.34	87.53	86.0	84.52	85.83	84.86
Water	95.3	95.29	91.54	94.41	95.79	95.33

---

Dataset

VDD (Varied Drone Dataset) is a semantic segmentation benchmark for drone imagery spanning varied altitudes, viewpoints, and scenes (urban, rural, and natural), captured at a uniform native resolution of 4000x3000.

Classes

• Other
• Wall
• Road
• Vegetation
• Vehicle
• Roof
• Water

---

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-large-vdd",
    filename="cabinet_best.pth"
)

model = CABiNet(n_classes=7, mode="large")
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	1000
Image size	1024
Batch size	4
Dataset	VDD (converted images/+masks/ format)
Framework	CABiNet custom PyTorch trainer
cls_pw (class weighting)	0.5

---

Official Resources

• VDD Semantic Segmentation Model Zoo: https://huggingface.co/collections/dronefreak/vdd-semantic-segmentation-model-zoo
• CABiNet repository: https://github.com/dronefreak/CABiNet
• CABiNet Paper: https://arxiv.org/abs/2011.00993v2
• VDD Dataset (Hugging Face): https://huggingface.co/datasets/RussRobin/VDD
• VDD Repository: https://github.com/RussRobin/VDD
• VDD Paper (arXiv): https://arxiv.org/abs/2305.13608
• VDD Published (JVCIR): https://www.sciencedirect.com/science/article/pii/S1047320325000434
• 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:

• Small training set (280 images) — heavier augmentation (mosaic/mixup/copy-paste) offsets this during training, but rare-class generalization may still be limited
• Rare classes (Vehicle, Roof, Water) are underrepresented relative to Vegetation/Road/Wall
• Very high native resolution (4000x3000, uniform) downsampled to the eval imgsz — fine detail on small objects (e.g. vehicles at altitude) can be lost
• Varied altitude/viewpoint scenes (the dataset's defining trait) can shift the domain between training crops and a given inference image

---

Citation

Please cite the following:

@article{cai2025vdd,
  title={Vdd: Varied drone dataset for semantic segmentation},
  author={Cai, Wenxiao and Jin, Ke and Hou, Jinyan and Guo, Cong and Wu, Letian and Yang, Wankou},
  journal={Journal of Visual Communication and Image Representation},
  volume={109},
  pages={104429},
  year={2025},
  publisher={Elsevier}
}

@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}
}