Much Cheaper Ground Filtering

Groundiff is a paper I have discussed previously on my blog. The premise is relatively simple, a diffusion model that converts an DSM to a DTM, by treating the scene as a noise removal problem.

I trained a model to get 28cm RMSE on the EA tiles, and an MAE of 10cm. But I consistently ran into a problem, that the model was too expensive to run on a regular computer or laptop. By using model distillation, based on this paper "Fine-Tuning Image-Conditional Diffusion Models is Easier than You Think", I fine tuned the model to run on 1 step rather than 10 steps.

With this extra step, I have achieved a model which can run within QGIS, that takes 2 minutes to run on my laptop, and can produce an nDSM and DTM from a DSM raster from the Environment Agency LiDAR.

You can find the code here - https://github.com/JacobWeinbren/GrounDiff

As for the model...I am not exactly sure how to host it, but I am happy to email it to you (please, of course, do a virus check on it first!). It is an onnx file, so works on Mac and PC.