zju3dv/NeuralRecon

Health v2 · maintenance only

Flags: no_releases

Adoption not part of the score

2274 stars · 312 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

NeuralRecon is a deep learning framework for real-time 3D scene reconstruction from monocular video with known camera poses. It reconstructs dense, coherent 3D geometry by sequentially predicting sparse TSDF volumes for video fragments using a neural network with a learning-based fusion module.

Use cases

• reconstruct 3d scene geometry from monocular video in real time
• generate dense 3d surface models from iphone arkit captures
• run 3d reconstruction on scannet dataset with pretrained weights
• build coherent 3d meshes from video fragments using neural tsdf fusion
• evaluate 3d reconstruction accuracy on indoor scene datasets

When to choose

• you need real-time dense 3d reconstruction from a single camera with known poses
• you want coherent geometry across frames rather than per-frame depth maps
• you have arkit or similar camera pose data from a mobile device
• you need a pretrained model that generalizes to new indoor scenes

When to avoid

• you lack camera pose information for your video input
• you need textured or colored 3d reconstruction rather than geometry only
• you work outside linux or cannot set up the torchsparse dependency
• you need reconstruction of large outdoor scenes rather than indoor environments

Facets

library · maturity maintenance

computer-vision machine-learning deep-learning graphics simulation computer-vision machine-learning deep-learning graphics artificial-intelligence python 3d-reconstruction monocular-video tsdf real-time cvpr-2021 scannet arkit neural-networks 3d-vision surface-reconstruction linux gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="zju3dv/NeuralRecon")