half-potato/radiance_meshes

Training code for "Radiance Meshes for Volumetric Reconstruction".

What it solves

This project provides the training code for Radiance Meshes, a method for volumetric reconstruction. It allows users to create 3D representations of scenes from a set of images, supporting various datasets like mipnerf360 and zipnerf.

How it works

The system takes a set of images of a scene and trains a model to reconstruct the volume. It supports custom dataset conversion via a convert.py script and is compatible with the dataset format used by Gaussian Splatting (via Agisoft Metashape). The training process can be optimized for different hardware constraints, such as reducing the hashmap size for lower VRAM GPUs, or adjusting parameters like vertex budget and voxel size to create reduced models specifically for mobile web viewing.

Who it’s for

Researchers and developers working in 3D reconstruction, volumetric rendering, and computer vision who have access to an NVIDIA GPU (ideally with 24GB VRAM) and are running on Linux.

Highlights

  • Volumetric Reconstruction: Focused on creating detailed 3D scenes from images.
  • Web-Optimized Training: Specific parameters provided to reduce model size and vertex count for mobile viewing.
  • Flexible Dataset Support: Compatible with standard NeRF datasets and Gaussian Splatting formats.
  • Hardware Adaptability: Settings can be adjusted to accommodate GPUs with less than 24GB of VRAM.

Related

  • Project
    dkogan/mrcal

    mrcal is a generic toolkit for solving camera calibration and Structure-from-Motion problems, providing C and Python libraries and command-line tools.

  • Project
    scribeocr/scribe.js

    A JavaScript library for performing OCR and text extraction from images and PDFs, capable of creating searchable PDFs with invisible text layers.

  • Project
    realsee-developer/RealSee3D

    A large-scale multi-view RGB-D dataset of 10,000 indoor scenes designed for training AI models in 3D reconstruction and semantic segmentation.

  • Project
    Sompote/DINOV3-YOLOV12

    A hybrid object detection framework combining YOLOv12 and DINOv3 to provide superior accuracy and faster convergence, especially on small or complex datasets.