Street Gaussians without the label tax

What it does S3Gaussian models dynamic street scenes for autonomous driving using deformable 4D Gaussians. It decomposes scenes into static and dynamic components and renders novel views without relying on 3D bounding boxes or other manual annotations. The system trains in two stages, bootstrapping from a prior checkpoint to extend coverage beyond an initial frame clip.

The interesting bit The project replaces expensive 3D labels with a multi-resolution hexplane encoder that compresses spacetime into feature planes, plus a multi-head decoder that outputs deformed Gaussians. It is essentially a self-supervised layer cake built atop 4DGaussians and EmerNeRF.

Key highlights

• Fully self-supervised: no 3D bounding boxes or extra annotations required during optimization.
• Decomposes dynamic and static elements from unlabeled driving video.
• Accepted to ICRA 2026; builds on 4DGaussians and EmerNeRF codebases.
• Supports novel view synthesis and outputs separate dynamic/static RGB and depth renderings.
• Two-stage training handles longer sequences by warm-starting from prior checkpoints.

Caveats

• Training is staged: extending beyond an initial clip requires loading a prior checkpoint rather than training continuously.
• The README notes support only for dynamic32 and static32 data splits, which may limit out-of-the-box dataset compatibility.
• Quantitative claims of “superior” decomposition and rendering quality are stated but not substantiated with numbers in the README.

Verdict Worth a look if you research neural rendering or autonomous driving perception and want to sidestep annotation pipelines. Skip it if you need a turnkey, single-stage trainer for arbitrary video.