AMD-Powered 3D Gaussian Splatting for Autonomous Driving Scenes

2026-05-07 · Source: AMD ROCm Blogs · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Software Development & Engineering · Depth: Advanced, long

Summary

AMD has released a new fork of Street Gaussians (SG), an extension of 3D Gaussian Splatting (3DGS), optimized for efficient training on AMD GPUs using ROCm™. This implementation addresses the limitation of original 3DGS to static scenes by explicitly separating dynamic street scenes into static backgrounds and multiple dynamic foreground objects, each represented by 3D Gaussians. The new release migrates SG to the latest gsplat API (v1.5.3), simplifying the rendering pipeline and reducing Python-side overhead. The workflow, validated on AMD Instinct™ MI300X GPUs, includes installation on ROCm™, data pre-processing using the Waymo Open Dataset, training, and evaluation. New rendering features enable smoother video generation via pose interpolation and scene editing capabilities, such as adjusting vehicle speeds, trajectories, or removing/duplicating objects, to generate novel training data for autonomous driving systems. Performance benchmarks on Waymo scenes show PSNR values up to 37.25, SSIM up to 0.95, and LPIPS as low as 0.19.

Key takeaway

For AI Engineers developing autonomous driving systems, this AMD-optimized Street Gaussians implementation offers a robust platform for dynamic scene reconstruction and editing. You can leverage its explicit object separation to generate synthetic training data for critical scenarios, such as altered vehicle trajectories or speeds, which are difficult to collect in real-world conditions. This capability can significantly enhance the robustness and safety of your autonomous driving models.

Key insights

Street Gaussians on AMD GPUs enables dynamic 3D scene reconstruction and editing for autonomous driving.

Principles

• Decompose dynamic scenes into static and moving components.
• Use Fourier coefficients for time-varying object appearance.
• Combine L1, SSIM, depth, sky, and semantic losses for quality.

Method

The method involves building data pre-processing and Street Gaussian Docker files, downloading and extracting Waymo data, running a data processing pipeline, performing training, and then rendering and evaluating the trained model.

In practice

• Generate novel training data by editing vehicle trajectories.
• Interpolate training poses for smoother, higher framerate videos.
• Isolate moving vehicles or static backgrounds for analysis.

Topics

• 3D Gaussian Splatting
• Street Gaussians
• Autonomous Driving Scenes
• AMD Instinct MI300X
• ROCm Platform

Code references

• LightwheelAI/street-gaussians-ns
• ROCm/gsplat
• nerfstudio-project/gsplat
• silogen/ai-samples
• silogen/ai-samples

Best for: Machine Learning Engineer, Computer Vision Engineer, AI Engineer

Related on AIssential

• Local-GS: Accelerating 3D Gaussian Splatting via Tile-Local Warp Coherence — Local-GS enhances 3DGS rendering by aligning Gaussian processing with GPU warp execution for improved efficiency.
• Sensor2Sensor: Cross-Embodiment Sensor Conversion for Autonomous Driving — Sensor2Sensor converts monocular dashcam videos into multi-modal AV logs using a diffusion architecture and 4D Gaussian Splatting for data pairing.
• Deep Learning Approaches for 3D Medical Scene Completion: From Geometric Modeling to Generative Paradigms — The field of 3D medical scene completion has rapidly evolved from voxel-based methods to advanced generative and rendering-aware paradigms.
• What happens when you combine Nano Banana with Gaussian splat tours? — Combining Gaussian splats with generative AI enables interactive spatial reimagining of real-world 3D environments.
• EnerGS: Energy-Based Gaussian Splatting with Partial Geometric Priors — EnerGS uses an adaptive geometric energy field and decoupled optimization to improve 3DGS reconstruction in sparse, partially observed scenes.
• Less Gaussians, Texture More: 4K Feed-Forward Textured Splatting — LGTM uses textured Gaussian primitives to enable 4K novel view synthesis with fewer primitives.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AMD ROCm Blogs.