Adrian Ramlal

June 12th, 2026 – 12:00-1:00 pm, EC4-2101A

3D Gaussian Splatting (3DGS) has emerged as a leading method for photorealistic scene reconstruction from multi-view images, yet existing approaches treat reconstruction as a purely visual optimization problem, ignoring the physical laws that govern real-world scenes. This talk explores a bidirectional relationship between physics and vision within the 3DGS framework. In the forward direction, physical and geometric priors are introduced at each stage of the pipeline: upsampled point cloud initialization improves reconstruction quality without architectural changes, mesh-coupled Gaussian representations enable physics simulation of dynamic scenes, and differentiable rigid-body simulation provides trajectory supervision during object occlusion in 4D reconstruction. In the inverse direction, we examine how observed fracture behaviour in food materials can be used to recover latent material parameters via surrogate modelling and reinforcement learning, enabling novel simulation of physically plausible fracture dynamics. Together, these contributions demonstrate that neither physics nor vision alone is sufficient for faithful dynamic reconstruction and simulation, and that integrating the two disciplines yields measurable and qualitatively meaningful improvements across all stages of the pipeline.