The current Gaussian model for the resolution and efficiency provides the gradients w.r.t. panel parameters necessary for optimisation, however it results in large differences between the modelled panels and the physical panels, which is evident in the uncertainties in inferred variables.
We should adjust this model to better simulate the physical panels, whilst still providing the gradients necessary to optimise their parameters.
This will likely involve a broader, flatter distribution inside the panel, which drops off outside the panel. Ideally, the steepness of this drop-off and the flatness inside the panel should be adjustable. These could then be annealed during optimisation such that the xy & size of the panels are quickly optimised in the first half of training, and then the z position is adjusted in the latter half.
The current Gaussian model for the resolution and efficiency provides the gradients w.r.t. panel parameters necessary for optimisation, however it results in large differences between the modelled panels and the physical panels, which is evident in the uncertainties in inferred variables. We should adjust this model to better simulate the physical panels, whilst still providing the gradients necessary to optimise their parameters. This will likely involve a broader, flatter distribution inside the panel, which drops off outside the panel. Ideally, the steepness of this drop-off and the flatness inside the panel should be adjustable. These could then be annealed during optimisation such that the xy & size of the panels are quickly optimised in the first half of training, and then the z position is adjusted in the latter half.