Compositional Scene

Compositional scene representation aims to model visual scenes as collections of individual objects and their relationships, enabling more robust and interpretable scene understanding. Current research focuses on developing neural scene representations, often employing Gaussian splatting, neural radiance fields, or scene graphs, to achieve efficient and high-quality novel view synthesis and scene manipulation, including editing individual objects via language instructions. These advancements are significant for applications like augmented and virtual reality, image editing, and robotics, by allowing for more flexible and intuitive interaction with 3D environments. The ability to disentangle object properties from viewpoint and scene context is a key challenge driving ongoing research.