deeperfly

Markerless 3D pose estimation of tethered Drosophila from a multi-camera rig. deeperfly estimates camera parameters and 2D/3D keypoint locations from behavioral recordings through one linear pipeline: 2D pose → bundle adjustment → triangulation → visualization.

It is both a command-line tool and a Python library, and a modern rewrite of DeepFly3D, DeepFly2D and PyBundleAdjustment.

Install

uv tool install git+https://github.com/NeLy-EPFL/deeperfly --python 3.13 --torch-backend=auto

Run

deeperfly doctor                                             # check the install
deeperfly init config.toml                                   # write a config (edit if needed)
deeperfly run examples/data/ -c config.toml                  # 2D -> 3D -> video
deeperfly inspect examples/data/deeperfly_outputs/results.h5 # summarize the result

Where to go next

• Getting started — run the bundled example end to end, from install to a rendered 3D video.
• CLI usage — every command and flag.
• Writing configs — the config.toml, section by section.
• Library API — use the pipeline, bundle adjustment and geometry from Python.
• How it works — the pipeline stage by stage, plus the conventions the whole package shares.
• Reference — the complete configuration, output format and library API.

License

GPL-3.0-only. See LICENSE.