chamud/3dgs-101

3DGS-101 is a 10-chapter self-study textbook on a 3D reconstruction technique called 3D Gaussian Splatting, written as a story rather than a reference manual. The goal is to take someone who already knows a bit of Python and linear algebra all the way to building a floater-aware generative scene completion pipeline. Each chapter is designed as roughly a two-hour study session, and they must be read in order since each one builds on the previous. 3D Gaussian Splatting is a method for turning a set of photographs of a scene into a 3D representation you can render from any angle. The textbook explains the full pipeline: starting from how cameras turn 3D points into 2D pixels, moving through how structure-from-motion tools estimate camera positions from photos, then explaining what a Gaussian splat is mathematically, how it gets projected onto a 2D screen, how many splats are combined to produce a full image, and how the whole system is trained by comparing rendered images to real photos and adjusting the splats accordingly. Later chapters cover a common problem called floaters, which are spurious splats that appear in under-observed parts of a scene, and how to read and extend real research codebases. Prerequisites are listed honestly: Python with NumPy, basic linear algebra (matrix multiplication, eigenvectors, inverses), and enough calculus to know what a derivative means. A machine with a GPU is needed for chapters 7 and 8, which involve training a scene from your own photos. Prior knowledge of graphics programming, CUDA, or neural radiance fields is not required. The recommended reading environment is VS Code with the Markdown Preview extension active, because diagrams are written in Mermaid notation and render live in that view. An Excalidraw extension handles the hand-drawn figures. Each chapter includes short Python coding exercises to make the concepts concrete. The final capstone chapter brings everything together: fusing a generative prior into the training loop to fill in parts of a scene that the original photos did not capture well.