Optimization and parallelization of a texture synthesis method for the VFX industry focusing on moving fluids. Parallel k-d trees construction for parallel closest point find on mesh with warp-level synchronized traversal.Parallelization with CUDA, OpenMP. Deterministic implementation.Unit tests with catch2 and built-in instrumentation library to validate each pipeline stage (C++/Python).Multiple independant sub-projects to avoid strong dependencies with CMake.Profilers and 3rd party libraries used: OpenCV, Thrust, SideFX Houdini, NVTX, Intel VTune, and NVIDIA Nsight Compute/Systems.

Exploration of the use of a generalist 3D reconstruction AI based on depth inference applied to a specific, planar configuration of 4x4 cameras (ReVeRY project in collaboration with URCA and XD Productions). Hyperparameters tuning (including camera intrinsics) and training of the neural network (MVSNet) in Docker containers, and adaptation of the dataset to be able to use the neural network with the dataset provided (Python).