S3Gaussian

Why we included this project

Autonomous driving teams that need to reconstruct dynamic street scenes from camera footage will get a lot out of this codebase. S3Gaussian models moving vehicles and static background with 3D Gaussian splatting, and it does the whole thing without 3D bounding box labels, using a self-supervised objective instead. A multi-resolution hexplane encoder compresses the 4D scene into feature planes, and a multi-head Gaussian decoder turns those into deformed 4D Gaussians, which is how the model keeps dynamic and static content apart while still rendering clean novel views. Since it builds on earlier work like 4D Gaussians and EmerNeRF, it is a solid starting point for reproducing the paper or pushing self-supervised street reconstruction further. Just know it is research code, not a plug-and-play product: you will need to prepare datasets and run the training pipeline yourself.