This project provides a lightweight segmentation model for detecting railroads on 50 cm resolution ortophotos of Slovenia, available from the official eProstor Geospatial Portal.
The model may be useful for image based localization and for educational purposes. It was trained on a single rtx 3090 without extensive parameter tuning, should be an interesting baseline for testing.
- Input size: 1024x1024 RGB orthoimage tiles
- Output: 1024x1024 probability tile
- Performance on edge devices: ~4.5 seconds per inference on a Raspberry Pi 5 (using ONNX runtime)
Below is an example showing the segmentation output:
Left: Input tile from orthophoto. Right: Railroad segmentation.
model.onnx– ONNX export of the trained modelinference_example.ipynb– Jupyter notebook demonstrating ONNX inference
- Improve dataset: The training dataset is currently quite small, not utilizing the whole country. The labeling also has some room for improvement
- Smaller and larger models: 512x512 model runs in 1.2s, might be interesting to go with a bigger backbone and compare the accuracy.