This project explores semantic segmentation of runway areas from aerial images to enable automated landing of rotorcraft on landing platforms. We experiment with a range of models—including both well-established architectures and novel attention-based designs—to evaluate their performance in real-world deployment conditions.
I’ve added a super–lightweight control module that uses the predicted runway segmentation mask to compute roll and yaw commands in under 4 ms. Combined with the segmentation inference time (~6.5 ms), the full perception‑to‑control pipeline now completes in just 10.6 ms. check out in my kaggle notebook
- The dataset consists of RGB images of runways captured from a rotorcraft's perspective.
- The ground truth masks are binary segmentation maps with white regions representing the runway, and black indicating background.
- The dataset includes images captured in diverse weather conditions (clear, cloudy) and during both day and night, making the segmentation task more challenging and realistic.
- Sample Image Structure:
Input Image→ Raw RGB aerial viewGround Truth Mask→ White (runway) on Black background
You can explore the dataset and learn more here.
To compare the efficiency and effectiveness of attention-based segmentation models (like MobileViT and Axial Attention) with traditional models (UNet, DeepLabV3+) for accurate and real-time runway detection.
| 🎯 Model | 🔍 Test Loss | ⏱ Inference (ms/img) | 📊 Params (Million) | 📈 Epochs to Converge |
|---|---|---|---|---|
| UNet (ResNet-34) | 0.0115 | 6.08 | 24.44 | 33 |
| DeepLabV3+ (ResNet-18) | 0.0095 | 4.94 | 12.33 | 21 |
| DeepLabV3+ (MobileNetV2) | 0.0086 | 6.58 | 4.38 | 24 |
| UNet (EfficientNet-B0) | 0.0089 | 12.90 | 6.30 | 34 |
| Axial-UNet (Experimental) | 0.0147 | 7.19 | 20.65 | 44 |
| MobileViT-UNet (Experimental) | 0.0125 | 7.74 | 11.89 | 52 |
| MobileViT-UNet Lite (Experimental) | 0.0183 | 3.79 | 2.89 | 61 |
These evaluations were conducted under identical training conditions:
- Hardware: NVIDIA Tesla P100 GPU
- Batch size: 64
- Image dimensions: 512×512
- Consistent train-test split
- Learning Rate : 1e-3
- Optimizer: Adam
- Loss: Binary Cross-Entropy with Logits (
nn.BCEWithLogitsLoss) - Random seed: 33
- DeepLabV3+ with MobileNetV2 achieves the best accuracy-efficiency trade-off.
- Attention-based models (Axial, MobileViT) show potential but require more epochs and have higher test losses—suggesting room for optimization.
- MobileViT-UNet Lite, despite being the lightest, performed the low in accuracy but is suitable for low-latency environments.
Below is a comparison of segmentation outputs for a sample image:
| 🎯 Model | 🔍 Results |
|---|---|
| UNet (ResNet-34) |  |
| DeepLabV3+ (R18) |  |
| DeepLabV3+ (MNV2) |  |
| UNet (EffNet-B0) |  |
| Axial-UNet |  |
| MobileViT-UNet Lite |  |
📁 View the complete notebook on Kaggle:
👉 Runway Platform Detection – Kaggle Notebook
- Incorporating temporal consistency for videos using ConvLSTM/Transformer-based approaches.
- Deployment and benchmarking on real drone edge devices (e.g., NVIDIA Jetson Nano, Raspberry Pi).
Soham Umbare
IIIT Raichur
📧 [email protected]
⭐ If you find this work interesting, consider giving it a star on Kaggle & GitHub!
🧑💻 Happy Experimenting! 🔬