zju3dv/Murre

What it solves

Murre addresses the challenge of high-quality multi-view 3D reconstruction from images. It improves upon standard monocular depth estimation by guiding the process with sparse depth data from Structure-from-Motion (SfM), resulting in more accurate and consistent 3D reconstructions.

How it works

The system operates in three main stages:

1. SfM Parsing: It parses COLMAP results to extract sparse depth maps, camera intrinsics, and poses.
2. SfM-guided Depth Estimation: A monocular depth estimation model (based on Marigold) is used to predict dense depth maps, using the sparse SfM data as a guide to ensure geometric consistency across multiple views.
3. TSDF Fusion: The resulting dense depth maps and camera parameters are fused using Truncated Signed Distance Function (TSDF) fusion to create a final 3D reconstruction.

Who it’s for

Researchers and developers working in computer vision, 3D scene reconstruction, and spatial AI who need high-precision dense 3D models from a set of images.

Highlights

• SfM Integration: Combines sparse, reliable SfM points with dense monocular depth predictions.
• Robust Filtering: Includes thresholds for reprojection error and minimum co-visible views to remove noisy depth estimates.
• Flexible Resolution: Allows users to trade off between inference speed and reconstruction precision via processing resolution settings.
• Scene Adaptability: Supports both indoor and outdoor scenes through adjustable depth clipping parameters.