Skip to primary content
SMS
Shah Mohammed SeamanSoftware Engineer & AI Researcher
Back to All Systems
AI & Computer Vision • 2024Completed

3D Gaussian Splatting Desktop Engine

High-performance desktop 3DGS renderer with dual CUDA and CPU fallback pipelines

Rendering Pipeline
CUDA + CPU Fallback
Supported Formats
Inria / Nerfstudio PLY
Harmonics Order
Degree 0 to 3 (f_rest_44)
Controls
Real-time Scale & Opacity
01 / EXECUTIVE OVERVIEW

Problem Context & System Scope

A minimal, high-performance desktop application designed to inspect, navigate, and render standard 3D Gaussian Splatting (.ply) scenes from Inria GraphDeco, Nerfstudio, Postshot, and Scan Studio.

Features a dual-pipeline architecture: utilizing gsplat GPU rasterization on CUDA-enabled hardware, while seamlessly providing a high-fidelity CPU projection pipeline with depth sorting and Spherical Harmonics evaluation when GPUs are unavailable.

Equipped with interactive orbit camera controls, scale multiplier, opacity thresholding, and diagnostic hardware runtime inspection.

02 / SYSTEMS ARCHITECTURE

Technical Breakdown & Pipeline Engineering

Pre-flight Hardware Diagnostic & Dual Pipeline

Dynamic runtime evaluation ensuring zero crashes.

  • Probes CUDA availability, PyTorch runtime capabilities, and gsplat binary bindings on launch.
  • Dynamically routes rasterization between gsplat GPU kernels and an optimized CPU depth-sorted projection fallback.

Standard Pretrained 3DGS PLY Parser

Strict specification compliance for Gaussian attributes.

  • Parses means (x, y, z), exponential scale activations (scale_0..2), and normalized wxyz quaternions.
  • Sigmoid activation handling for opacity values.
  • Evaluates complete Spherical Harmonics coefficients (DC terms and higher-order rest features f_rest_0..44).

Interactive Orbit Camera & Viewport Controls

Smooth 6-DOF navigation with real-time parameter tuning.

  • Low-latency orbit, middle-click pan, smooth zoom, and auto-fit bounding box calculation.
  • Real-time UI sliders adjusting splat scale multipliers (0.1x to 3.0x) and opacity limits.
03 / KEY CHALLENGES & SOLUTIONS

Engineering Hardships & Resolution

Challenge:
Preventing application failure on systems without modern NVIDIA hardware or mismatched CUDA toolkit versions.
Engineering Solution:

Implemented an isolated hardware probing routine that catches CUDA initialization exceptions and gracefully initializes a CPU-based spherical harmonics renderer.

04 / IMPACT & VALIDATION

Quantitative Results & Impact

  • Enables seamless local inspection of photorealistic 3DGS scenes directly from Python/Qt desktop environments.
  • 100% compliant with industry-standard PLY coordinate systems and quaternion conventions.
05 / APPLIED TOOLCHAIN

Technologies & Frameworks

PySide6 (Qt)PyTorchgsplat LibraryCUDASpherical HarmonicsPython 3.11