A Lightweight Dynamic Convolution Network with Attention Gates and Deep Supervision for UAV Semantic Segmentation
Real-time anti-UAV semantic segmentation with only 1.66M parameters.
Upload your own UAV image → DAS-Net detects it in real-time. 👉 https://huggingface.co/spaces/niceyj/DAS-Net-Demo
- Lightweight: 1.66 M parameters (14.7× smaller than ResNet-34 UNet)
- Accurate: Matches UNet (mean test mIoU 0.6780 vs. 0.6760) on a 174,008-image test set
- Real-time: 113 FPS on NVIDIA A6000 GPU (FP32, batch size 1, 512×512)
- Statistically validated: Three-seed Monte Carlo cross-validation (p = 0.045, Cohen's d = 1.74)
- Anti-UAV ready: Designed for resource-constrained UAV surveillance platforms
- 🎮 Try It Live
- Overview
- Visual Results
- Architecture
- Quantitative Results
- Installation
- Usage
- Citation
- License
- Acknowledgements
DAS-Net (Dynamic Attention-Supervised Network) extends the lightweight ThinDyUNet baseline with three architectural improvements for UAV semantic segmentation:
- Symmetric Dynamic Convolution: applied to both encoder and decoder paths (vs. encoder-only in ThinDyUNet)
- Attention Gates: filter encoder features at skip connections before concatenation with decoder features
- Deep Supervision: auxiliary loss heads at the last three decoder stages (λ = 0.4) — discarded at inference
The result is a 1.66 M-parameter model that achieves UNet-level accuracy at 14.7× fewer parameters and runs at real-time speeds on workstation GPUs.
Detection examples on three test scenes (top → bottom):
- Top row: Clear visible-light (VL) scene
- Middle row: Infrared (IR) scene with cluttered background
- Bottom row: IR scene with small, distant UAV target
Columns (left → right): Input image, Ground-truth mask, ThinDyUNet, DAS-Net (Ours), MobileUNet, UNet, PAN, PSPNet.
Observations: DAS-Net produces masks consistently closer to the ground truth than other models, particularly for small or partially occluded targets where ThinDyUNet often misses detection. The attention gates effectively suppress background clutter while the symmetric DyConv decoder restores fine-grained UAV boundaries.
Sample inputs: (a) visible light (VL), (b) IR under clear weather, (c) IR under foggy conditions. (d–f) Corresponding ground-truth segmentation masks. Note the small UAV target relative to the entire frame, which is a key challenge for accurate segmentation.
Trade-off between model size and segmentation accuracy: DAS-Net achieves the highest mIoU among lightweight models (≤ 2 M parameters) and is competitive with UNet (24.4 M) while using 14.7× fewer parameters, placing it strongly on the Pareto frontier.
DAS-Net architecture overview: Symmetric encoder-decoder with DyConvBlocks (K=2 parallel kernels) in both paths, attention gates filtering all skip connections, and deep supervision auxiliary heads attached to the last three decoder stages (used only during training).
| Model | Encoder | Decoder | Attention Gate | Deep Supervision | Params (M) |
|---|---|---|---|---|---|
| ThinDyUNet (baseline) | DyConvBlock | MultiCNNBlock | — | — | 1.34 |
| FullDyUNet | DyConvBlock | DyConvBlock | — | — | 1.63 |
| DeepSupDyUNet | DyConvBlock | MultiCNNBlock | — | ✓ (λ=0.4) | 1.34 |
| DAS-Net (Ours) | DyConvBlock | DyConvBlock | ✓ | ✓ (λ=0.4) | 1.66 |
| Model | Params (M) | Precision | Recall | Dice | mIoU | A6000 (ms) | FPS |
|---|---|---|---|---|---|---|---|
| ThinDyUNet | 1.34 | 0.9308 | 0.6449 | 0.6766 | 0.6115 | 8.29 | 120.6 |
| PSPNet | 21.4 | 0.8554 | 0.6685 | 0.6768 | 0.6094 | 11.46 | 87.3 |
| MobileUNet | 6.0 | 0.9248 | 0.6593 | 0.6790 | 0.6185 | 10.23 | 97.7 |
| PAN | 21.4 | 0.9055 | 0.7047 | 0.7045 | 0.6438 | 11.16 | 89.6 |
| FullDyUNet | 1.63 | 0.8816 | 0.6994 | 0.7159 | 0.6451 | 9.04 | 110.7 |
| DeepSupDyUNet | 1.34 | 0.9268 | 0.7063 | 0.7222 | 0.6627 | 8.26 | 121.0 |
| UNet | 24.4 | 0.9004 | 0.7333 | 0.7413 | 0.6760 | 9.93 | 100.7 |
| DAS-Net (Ours) | 1.66 | 0.8422 | 0.7643 | 0.7509 | 0.6780 | 8.83 | 113.2 |
DAS-Net matches UNet's mIoU (0.6780 vs. 0.6760) with 14.7× fewer parameters and achieves the highest mIoU among lightweight models (≤ 2 M parameters).
DAS-Net converges smoothly and achieves the highest validation mIoU by the end of training, with the lowest validation loss among all compared models.
- Python 3.8+
- PyTorch 2.0+
- CUDA 11.x or 12.x
- NVIDIA GPU (8+ GB VRAM recommended for training)
# Clone repository
git clone https://github.com/niceyoungjae/DAS-net.git
cd DAS-net
# Create conda environment (recommended)
conda create -n dasnet python=3.10 -y
conda activate dasnet
# Install dependencies
pip install -r requirements.txtSee docs/INSTALLATION.md for detailed setup including dataset preparation.
We use the UAV semantic segmentation dataset proposed in Kim & Jang, Appl. Sci. 2025.
- Reference paper: https://doi.org/10.3390/app15137183
- Public source: https://github.com/SCKIMOSU/uav
- Size: 605,045 paired VL + IR images (1920×1080)
- Split: 304,677 train / 126,360 val / 174,008 test
For full dataset access, please refer to the original publication or contact the dataset authors at sckim7@kookmin.ac.kr.
Each model has its own dedicated training script:
# Train DAS-Net
python scripts/train_dasnet.py --config configs/config_dasnet.yaml
# Train ablation variants
python scripts/train_thindyunet.py --config configs/config_thindyunet.yaml
python scripts/train_fulldyunet.py --config configs/config_fulldyunet.yaml
python scripts/train_deepsupdyunet.py --config configs/config_deepsupdyunet.yaml
# Train external baselines
python scripts/train_unet.py --config configs/config_unet.yaml
python scripts/train_mobileunet.py --config configs/config_mobileunet.yaml
python scripts/train_pan.py --config configs/config_pan.yaml
python scripts/train_pspnet.py --config configs/config_pspnet.yaml# Evaluate DAS-Net (loads checkpoint from configs)
python scripts/test_dasnet.py --config configs/config_dasnet.yaml
# Or use the generic eval script
python scripts/eval.py --model dasnet --checkpoint <path-to-best.pth>See docs/USAGE.md for advanced options including multi-seed training and custom inference.
DAS-net/
├── README.md ← You are here
├── LICENSE ← MIT License
├── requirements.txt ← Python dependencies
├── CITATION.cff ← How to cite
├── configs/ ← YAML configs for 8 models
│ ├── config_dasnet.yaml
│ ├── config_thindyunet.yaml
│ └── ...
├── model/ ← Model definitions
│ ├── DASNet.py
│ ├── ThinDyUNet.py
│ ├── FullDyUNet.py
│ ├── DeepSupDyUNet.py
│ └── _base.py
├── dataset/ ← Data loader
├── scripts/ ← Training & evaluation
│ ├── train_*.py ← Per-model training (8 total)
│ ├── test_*.py ← Per-model evaluation (8 total)
│ ├── trainer.py ← Common training utilities
│ └── eval.py ← Generic evaluation
├── utils/ ← Common utilities (early stop, model save)
├── docs/ ← Detailed documentation
└── figures/ ← Paper figures (PNG)
If you find this work useful, please cite our paper:
@article{kim2026dasnet,
title = {DAS-Net: A Lightweight Dynamic Convolution Network with Attention Gates
and Deep Supervision for UAV Semantic Segmentation},
author = {Kim, Young Jae and Kim, Sang-Chul},
journal = {Applied Sciences},
year = {2026},
note = {Under review (Major Revision Submitted, applsci-4300704)}
}Also consider citing the underlying dataset:
@article{kim2025dataset,
title = {A Semantic Segmentation Dataset and Real-Time Localization Model
for Anti-UAV Applications},
author = {Kim, Sang-Chul and Jang, Yeong Min},
journal = {Applied Sciences},
volume = {15},
pages = {7183},
year = {2025},
doi = {10.3390/app15137183}
}This project is licensed under the MIT License — see LICENSE for details.
- Funding: This work was supported by the Korea Research Institute for Defense Technology Planning and Advancement (KRIT) grant funded by the Korean government [DAPA] under Grant KRIT-CT-23-041 (LiDAR/RADAR-Supported Edge AI-based Highly Reliable IR/UV FSO/OCC Specialized Research Laboratory, 2024).
- Dataset: We thank the authors of Kim & Jang, 2025 for the UAV semantic segmentation dataset.
- Implementations: External baselines (UNet, MobileUNet, PAN, PSPNet) are implemented via
segmentation_models.pytorch.
- Young Jae Kim (First Author): niceyj16@kookmin.ac.kr
- Sang-Chul Kim (Corresponding Author): sckim7@kookmin.ac.kr
- Affiliation: Department of AI·SW / School of Computer Science, Kookmin University, Seoul, Republic of Korea
For questions, please open a GitHub Issue.
⭐ Star this repo if you find it useful!