Finally found the bug which was killing the performance: Wrong boundary conditions in reconstruction and explicit casting from uint8_t to int / float using static_cast
Used NN without deconvolution kernel and GPUCF with deconvolution (standard)
Newest insight
- CNN works better for initial layers than fully connected
- 3D input works much better for deconvolution than 2D input. Current input size: 5x11x11 (row x pad x time)
Now comparing the reconstruction quality.
- Similar results for non-distorted and distorted data
- NN appears a bit worse still by the official metrics (e.g. chi2) but I don't think its correct (see plots)

Checked some individual matched tracks, their Chi2/Ncl and the resulting MSE error in Z direction
The network is consistently better in the MSE even if the chi2/Ncl is way worse for the NN

This means, the resulting differences in Chi2/NCl is either influenced by a calibration or by setting the split-flags (currently not set using the NN)