Skip to content

🏷️ Release v0.1.0 — “First Light”

Choose a tag to compare

View all tags
@abhaskumarsinha abhaskumarsinha released this 07 Nov 04:06
· 5 commits to main since this release
v0.1.0
4339223
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

🏷️ Release v0.1.0 — “First Light”

Date: November 2025
Author: Abhas Kumar Sinha
Repository: Gaussian LiteSplat

✨ Overview

Gaussian LiteSplat v0.1.0 marks the first public release of the project —
a fully functional, minimal implementation of Gaussian Splatting built in Keras (Torch backend).

This release focuses on:

  • Portability (runs on CPU and lightweight GPUs)
  • Simplicity (no CUDA or heavy dependencies)
  • Transparency (clear and readable training & rendering pipeline)
  • Accessibility for research and education

🧩 Core Features

âś… COLMAP Importer

  • Supports importing 3D reconstructions from COLMAP
  • Compatible with camera models:
    • SIMPLE_PINHOLE
    • PINHOLE
    • SIMPLE_RADIAL (without k)

âś… Gaussian Training Pipeline

  • End-to-end training of 3D Gaussian primitives
  • Trainable RGB color (no SHs, simplified version)
  • CPU-compatible with optional Torch compilation
  • Easily adjustable Gaussian count and image resolution

âś… Rendering System

  • Differentiable rasterization via Gaussian accumulation
  • View-consistent reconstruction from multi-view data
  • Supports real-time visual monitoring and output export

âś… I/O Utilities

  • Import / Export of Gaussians (gaussians.json) and Cameras (cameras.json)
  • Reproducible configuration through command-line arguments

âś… Colab Support

  • Works out of the box on Google Colab
  • No extra installations or system setup required

âś… Notebooks & Scripts

  • Example Jupyter notebooks for import, training, and visualization
  • CLI script for automated training (scripts/train_colmap.py)
  • Benchmarking and projection experiment templates

đź“Š Demonstration

  • Dataset: TempleRing (Middlebury Multiview)
  • Training: 2200 Gaussians, 15 Cameras
  • Hardware: CPU-only (no CUDA)
  • Training Time: ~45 minutes
  • Output: Progressive Gaussian density convergence visible after a few epochs

🧠 Known Limitations

  • No spherical harmonics (RGB-only colors)
  • Limited to a few COLMAP camera models
  • Not optimized for high Gaussian counts (>10k)
  • Currently supports only Torch backend for Keras

🚀 Next Milestones (v0.2.0 Roadmap)

  • Add SH-based color expansion
  • Support for more COLMAP camera models
  • Optimized CPU/GPU kernel paths
  • Simple GUI visualizer for scene exploration
  • Integration of benchmark comparisons

🧩 v0.1.0 establishes the foundation of Gaussian LiteSplat —
a lightweight, accessible, and educational Gaussian Splatting sandbox for everyone.

Full Changelog: https://github.com/abhaskumarsinha/Gaussian-LiteSplat/commits/v0.1.0

Assets 2
Loading