Learning to Infer 3D Shape Programs with Differentiable Renderer
This work addresses the challenge of representing and reasoning with high-level regularities in 3D shapes for AI and cognitive science, offering an incremental improvement over prior methods.
The authors tackled the problem of learning 3D shape programs for objects like tables and chairs by proposing a differentiable executor that improves faithfulness and controllability in interpreting shape programs, requiring no training and enhancing sample efficiency. Preliminary experiments demonstrated these advantages, encouraging further exploration in machine reasoning.
Given everyday artifacts, such as tables and chairs, humans recognize high-level regularities within them, such as the symmetries of a table, the repetition of its legs, while possessing low-level priors of their geometries, e.g., surfaces are smooth and edges are sharp. This kind of knowledge constitutes an important part of human perceptual understanding and reasoning. Representations of and how to reason in such knowledge, and the acquisition thereof, are still open questions in artificial intelligence (AI) and cognitive science. Building on the previous proposal of the \emph{3D shape programs} representation alone with the accompanying neural generator and executor from \citet{tian2019learning}, we propose an analytical yet differentiable executor that is more faithful and controllable in interpreting shape programs (particularly in extrapolation) and more sample efficient (requires no training). These facilitate the generator's learning when ground truth programs are not available, and should be especially useful when new shape-program components are enrolled either by human designers or -- in the context of library learning -- algorithms themselves. Preliminary experiments on using it for adaptation illustrate the aforesaid advantages of the proposed module, encouraging similar methods being explored in building machines that learn to reason with the kind of knowledge described above, and even learn this knowledge itself.