GilesStrong
commented
2 years ago Notes from a discussion with @tdorigo and @vischia proposing a statistic based on the difference between the mean predicted X0s of all voxels and that of the bottom 10 X0 predictions. I.e. if the distance increases, then it is likely that a small amount of dense material is present. The X0 predictions are also recalibrated based on the predictions of neighbouring voxels. Maybe @tdorigo can comment more.
In some cases, we are more interested in predicting the presence of a certain material, rather than accurately imaging the passive volume. In such cases and when we are using the model-inversion+PoCA, we need to convert a set of X0 predictions per voxel (or potentially a set of X0 predictions per muon) into a single number that can be fed into a sigmoid (if not already in [0,1]). Additionally, in order for such a prediction to be useable for detector optimisation, it must be differentiable. Further, since it is likely that such an inference algo. will be compared against the VoxelNet model, the algo. should not be too advanced, and ideally be as "classical" as possible. This would also help towards the X0-bias mitigation discussed in #94