To massively simplify a complex debate, there are 2 main schools of thought with respect to how to decrease future crime once a criminal has been apprehended [1]. The first is to make the punishment disincentivize future transgressions, and the second is to rehabilitate the criminal so that they won't make the decision to break the law again. Recently, society has been coming to terms with the fact that strict sentencing laws don't work in many situations (and often have unforeseen consequences, like the Revolving Door effect in American prisons [2]). In this situation, as much as we wish we could have directed the poor man to the nearest homeless shelter serving hot soup, it doesn't seem likely that taking away the threat of a 6 month prison stay would make him less likely to do that. The various hardships that are overindexed in the homeless population, from drug addiction to mental illness [3] (to perhaps a language barrier in this case) not only make them less likely to be able to find food consistently, but also make them less likely to respond to criminal disincentives.
It is a more enlightened philosophy of justice to acknowledge that saying 'you should have known better' won't make people know better. As the concept of morality progresses from a simple Hammurabian ear-for-an-ear justice game to a nuanced empirically-founded optimization problem, it will be interesting to see whether or not people continue to take pleasure in exerting it. That the latter problem tends to be incredibly engrossing to the technically literate means many of us will respond to this quite positively.
Postscript: If morality were a solved problem, would there be as many Effective Altruists?
[1] http://www.apa.org/monitor/julaug03/rehab.aspx
[2] http://www.pewtrusts.org/en/research-and-analysis/reports/00...
[3] https://www.researchgate.net/profile/Adriana_Foster/publicat... | relevant quote : estimation of mood disorders at "12% to 30% in various homeless samples".