what is gaussian_splatting_recipe fr?

gaussian_splatting_recipe is a Python pipeline that turns a drone video into a navigable 3D model using a technique called Gaussian Splatting, which represents a scene as millions of small colored blobs that can be viewed from any angle. The project is built specifically around video shot with DJI drones. The pipeline runs in a set order of steps. First, it pulls individual frames out of an input MP4 video file. It can optionally use an object detection model, Grounding DINO, to automatically mask out specific objects you name, such as a bench or a backpack, so they can be excluded or focused on separately. Next, it matches features between neighboring frames using two computer vision tools called SuperPoint and LightGlue, producing a 3D reconstruction of the camera's positions and the scene's structure in a format called COLMAP. A separate diagnostic script lets you visually check how well those feature matches lined up before continuing. The final training step takes that reconstruction and optimizes a Gaussian Splatting model from it, saving the result as a single PLY file. A small web based viewer, powered by the PlayCanvas engine, lets you open that PLY file directly in a browser and fly around the reconstructed scene by dragging and dropping it onto the page. Setting this up involves installing a specific combination of dependencies in order: a package called fVDB along with a matching CUDA build of PyTorch, the project's own Python requirements file, and a separate open source library called Hierarchical-Localization cloned from its own repository. This suggests the pipeline expects an Nvidia GPU with CUDA support rather than running purely on a CPU. Each script is run from the command line with paths to your video and output folders as arguments, in the order: extract frames and match features, optionally check the matches visually, then train and export the splat model. The README does not state a software license.