Gemini connector

Abstract Spectacular advances have been made in the field of machine vision over past decade. While this discipline is traditionally driven by geometric models, neural networks proven to be superior some applications and significantly expanded limits what possible. At same time, conventional graphic models describe relationship between images associated scene with textures light a physically realistic manner are an important part photogrammetry. Differential renderers combine these approaches enabling gradient-based optimization fixed structures graphics pipeline thus adapt learning process networks. This fusion formalized knowledge motivates idea modular differentiable renderer which physical statistical can recombined depending on use case. We therefore present Gemini Connector: initiative for development combination examine opportunities problems motivate extension rendering include underwater optics analysis deep sea images. Finally, we discuss cases, especially within Cross-Domain Fusion initiative.