Inverse Rendering
Image Credit: Ziang Liu- Close the sim2real gap by developing realistic sensor model that allows learning policies in simulation that can be ported to real world.
- Implemented FCNN using Flux.jl and CUDA.jl to learn scene geometry isosurface using signed distance function (SDF) values as input.
- Implemented sphere tracing and BRDF models, and used pixel L2 loss to optimize geometry pose and properties based on a single depth image.
