Material learning curves for real measurement data instead of synthetic data

[siebertdarius]

Hello Nvidia,

I generated the material learning curves as in Fig. 3 of you paper Learning Radio Environments by Differentiable Ray Tracing but instead of the synthetic dataset (as in Synthetic_Data.ipynb) i used the real dichasus-dc01 dataset (as in Learned_Materials.ipynb).

You clearly showed that the material properties can be learned by training from synthetic data, but unfortunatley the paper does not state if this also works for the real measurement data what I want to find out.

By adapting the plotting function for the training results I obtained the following curves for the real dataset.
As you can see, the material properties are way off in comparison to the synthetic data. Also, the loss stays at a certain minimum but fluctuates a lot for every next iteration/measurement datapoint. Thus I assume, the training found its rather unstable minimum.

Do you have the same observation when using the real measurement data?

I appreciate your comment on this.