FYI: WIP / research codebase.
Goal: provide an API for writing parameterized procedural models of environments by writing nodes and rules to be used in a scene grammar. We provide first-class support for:
- Forward sampling to generate lots of new scenes as collections of objects. These can then be rendered into images, forward simulated, etc.
- Scene parsing: given an observed environment, find maximum-a-posteriori program traces of the procedural model that reproduce it.
- Parameter estimation: Given a dataset of observed environments (where each observation is a collection of observed objects with observed poses, but without annotated relationships or structure), find optimal grammar parameters to best match that dataset.
For now, refer to this paper for the most up-to-date description of the grammar and parsing strategy. Refer to the unit tests and examples for usage. I promise more complete docs are on their way...