facebookresearch/map-anything

Health v2 · maintenance only

Adoption not part of the score

3682 stars · 283 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

MapAnything is an open-source research framework from Meta and CMU for universal feed-forward metric 3D reconstruction using an end-to-end trained transformer. It supports over 12 reconstruction tasks including multi-view stereo, structure-from-motion, monocular depth estimation, registration, and depth completion, with a complete stack for data processing, training, inference, and profiling.

Use cases

• reconstruct 3D scenes from multiple images
• estimate metric depth from a single image
• run structure-from-motion without classical pipelines
• complete missing depth in partial depth maps
• register point clouds or images using poses and calibration
• build 3D perception for robotics from camera feeds
• convert images to 3D geometry

When to choose

• you need a single feed-forward model covering many 3D vision tasks
• you want metric-scale 3D reconstruction from images with optional calibration or pose inputs
• you need a research framework with training and inference code included
• you are working on robotics or 3D perception research

When to avoid

• you need a lightweight production-ready 3D pipeline with minimal dependencies
• you require real-time reconstruction on CPU-only hardware
• you need classical photogrammetry with fine-grained bundle adjustment control
• your project is not in Python or lacks GPU support

Facets

library · maturity active

machine-learning deep-learning computer-vision image-processing simulation computer-vision machine-learning robotics artificial-intelligence python windows 3d-reconstruction multi-view-stereo depth-estimation sfm transformer metric-3d depth-completion calibration gpu linux macos

Member repositories

For agents

markdown · JSON · MCP: product_card(name="facebookresearch/map-anything")