Implement 2-pass trace integration

[kosack]

We should have an implementation of the 2-pass trace-integration (using timing info) that is described e.g. in Holder et al, 2006. The result is a higher signal-to-noise image, and the ability to apply lower cleaning thresholds in subsequent steps.

The basic procedure is:

1. use a wide (20ns) trace integration
2. apply a 1st stage image cleaning (tailcuts)
3. calculate the time gradient of the resulting pixels
4. use the time-gradient to predict the position of the pulses to better accuracy
5. re-integrate with a 10ns window using the information from step 4

This is not so far different from the NeighborPeakIntegrator with some extra time information, and would be useful for comparing our results with MARS and EventDisplay.

[watsonjj]

This is very similar to a charge extraction technique that Heide Costantini implemented in read_hess. I have been intending to re-implement it into ctapipe for a long time. But I would be happy for someone else to do it if I don't get to it in time.

[kosack]

@watsonjj yes, I think it's identical in fact (read_hess, MARS, and VERITAS all implemented it). If you have time, that would be great!

[GernotMaier]

Some points which we learnt using the double pass method:

• use a robust fitter for the determination of the time gradient (a single outlier might otherwise pull the direction of the image axis into the completely wrong direction)
• pass 1 needs reasonable high cleaning levels to determine a good image long axis. But this means that you miss some of the smallest images. Don't discard them - simply use as integration time the typical (average) pulse arrival time
• the assumption is obviously that some calibration has been applied to the pixel arrival times (if necessary for both high and low gain chain)

[kosack]

Is there a reason why you would do a fit to determine the time gradient rather than just calculating it with a perhaps weighted average? It's essentially the same thing as calculating the moments. You can get rid of outliers the same way as for the moment calculation.

[dneise]

2-pass trace-integration (using timing info) that is described e.g. in Holder et al, 2006.

For completeness, I think you mean this? https://arxiv.org/pdf/astro-ph/0604119.pdf page 9 ff

[dneise]

Somebody working on this already?

[kosack]

just noticed this isn't linked to #1215

[HealthyPear]

@kosack @watsonjj

We can now review this discussion in light of what has been merged from #1215 , namely did that PR take into account all what has been discussed here? If yes, I think we can close this issue.

I think that everything is there, aside from improving or adding more fitting algorithms for the timing information (this can be done later I guess).

[maxnoe]

I glimpsed over the two-pass integration when doing #1336, there is a lot to be improved still. I'll make an issue with things that catched my eye.