Home > Visual Scene Grammars

Develop a modular grammar for dynamic scene representations.

Start by creating a simple grammar generator and validator. Establish a rendering pipeline in either Blender or Gazebo to visualize scenes. Explore the application of grammatical evolution for program synthesis, and implement an outer-inner-optimizer framework for mixed fine-tuning and grammatical hyperparameter selection. Expand the module set and conduct probabilistic inferences needed for Bayesian state estimation.

Utilize pre-trained neural networks for building and fine-tune as needed. Implement online inference methods to enhance real-time capabilities. Draw inspiration from the works of modular neural ordinary differential equations and modular meta-learning.

This project aims to advance state estimation and learning in observers for dynamic scenes, with an overarching goal of achieving a emergent modular grammar system for scene representation.


Data generation, ground truthing, and first ideas 23 August, 2023

Program Synthesis for Concatenation Halloween, 2023