zju3dv/EasyVolcap

What it solves

EasyVolcap is a PyTorch library designed to accelerate research in neural volumetric video. It simplifies the complex process of volumetric video capturing, reconstruction, and rendering, providing a unified framework that reduces the boilerplate code needed to implement and test new volumetric video algorithms.

How it works

The library provides a modular architecture that separates the pipeline into samplers, networks, and renderers. It supports both a unified NeRF-like pipeline (where components can be swapped via configuration files) and a fully customized network construction approach. The framework handles data loading, camera pose correction, and image undistortion, allowing researchers to focus on the core network logic. It also includes a WebSocket-based server-side rendering system to decouple rendering from viewing.

Who it’s for

It is intended for researchers and developers working on 4D view synthesis, neural radiance fields (NeRF), 3D Gaussian Splatting (3DGS), and other volumetric video reconstruction technologies.

Highlights

• Multi-Algorithm Support: Includes implementations of Instant-NGP+T, 3DGS+T, and ENeRFi.
• Flexible Customization: Allows users to register custom modules using a simple decorator and swap components via YAML configs.
• Integrated Tooling: Provides a GUI viewer, rendering servers, and scripts for extracting geometry (point clouds) for initialization.
• Extensible Design: Encourages a forking workflow for new projects, as demonstrated by the 4K4D project.