Ah, so they wrote a DSL just for doing Bayesian inverse-vision models, and apparently they've now got it to accuracy rates competitive with most other major vision methods?

Good job!