• [2026-06] 🎉 MAMMA being presented at CVPR 2026
• [2026-06] Code released (inference + training)

git clone https://github.com/cuevhv/mamma.git
cd mamma

Full env + CUDA + weights setup: docs/INSTALL.md.

micromamba activate mamma          # or: conda activate mamma
python -m inference doctor         # verify env vars + weight paths

The pipeline is zero-config when weights live under data/.

Bundled 4-cam example, ~56 MB:

bash data/download_example.sh                                       # fetches videos to data/mamma_example/

python -m inference run \
  --cfg      configs/examples/presets/quick.yaml \
  --footage  data/mamma_example \
  --seq_name pushing_and_lifting_from_ground \
  --calib    configs/examples/calib/iphones_outdoors.yaml \
  --out-tag  demo -v

Outputs land under output/ma_*/demo/mamma_example/….

Prefer a browser UI? Run bash gui/scripts/dev.sh, open http://localhost:3000, and click Run demo. It's the same pipeline but friendlier UX!

ma_cap → ma_masks → ma_2d → ma_3d → ma_vis

Entry point: python -m inference run (source: inference/cli/run.py).

Three things are needed:

1. A calibration file (how to make one)
2. A folder with your sequence (how to set it up)
3. A preset — use a shipped one: configs/examples/presets/quick.yaml (~5 min smoke) or configs/examples/presets/full.yaml (full-frame). See docs/CONFIGS.md to modify or author your own.

Then:

python -m inference run \
  --cfg      <path/to/preset>.yaml \
  --footage  <path/to/footage> \
  --seq_name <seq_name> \
  --calib    <path/to/calib>.yaml \
  --out-tag  run01 -v

Alternative — iterate over many sequences in one invocation. A capture JSON enumerates sequences, cameras, and the calibration in one file; the runner walks them automatically:

python -m inference run \
  --cfg     <path/to/preset>.yaml \
  --capture <path/to/capture>.json \
  --out-tag run01 -v

This fork adds the following accuracy-neutral speed and memory improvements over upstream MAMMA. Mask/landmark outputs are unchanged (verified byte-identical where applicable).

In segmentation/configs/sam3.yaml (and sam3_default.yaml), under sam::

sam:
  offload_video_to_cpu: true   # keep loaded frames in CPU RAM (negligible speed cost)
  offload_state_to_cpu: true   # keep tracking state in CPU RAM (~10-15% slower tracking)

Both default to true. Set to false to trade RAM for speed when GPU headroom allows.

Very long 4K sequences: SAM3 still preloads the whole video tensor when fed a full MP4, so multi-thousand-frame 4K clips can exhaust system RAM. Limit the range with global.start_frame / global.end_frame (or --start / --end); a clipped range is extracted to frames and streamed safely.

Browser UI for submitting and inspecting runs. It uses the same mamma python env.

gui/scripts/dev.sh        # dev: Flask :8000 + Vite :3000 (auto-reload)
gui/scripts/prod.sh       # prod: single Flask process on :8000

Setup and deployment: gui/README.md.

The paper's released captures, evaluation data, and synthetic training data live on the MAMMA project page and require a free account.

1. Register at https://mamma.is.tue.mpg.de/ and confirm your email.

2. Either use the GUI's Pipeline assets panel (sign in once, click to download), or run the per-dataset shell scripts under data/:

   bash data/download_mamma_dance.sh --bachata --meta --pred --videos_crf24

Five dataset families ship: dance, multi-person, iPhone, eval, and synthetic. Per-dataset sizes, video encodings, and the full script flag surface live in docs/DATASETS.md.

Just running on your own footage? You don't need any of this — see Run the pipeline above.

.
├── inference/       runner, step builders, doctor CLI
├── capture/         ma_cap step
├── segmentation/    ma_masks step
├── landmarks/       ma_2d step
├── optimization/    ma_3d step
├── visualization/   ma_vis step
├── configs/         presets + capture manifests
├── data/            body models + weights + datasets (gitignored)
├── output/          run outputs (gitignored)
├── gui/             browser UI (Flask + React)
└── scripts/         smoke tests + utilities

• Release the evaluation scripts (2D landmark + benchmark evaluation) and the processed evaluation datasets.

@inproceedings{cuevas2026mamma,
  title     = {{MAMMA}: {Markerless Accurate Multi-person Motion Acquisition}},
  author    = {Cuevas Velasquez, Hanz and Yiannakidis, Anastasios and Shin, Soyong and Becherini, Giorgio and H{\"o}schle, Markus and Tesch, Joachim and Obersat, Taylor and Alexiadis, Tsvetelina and Halilaj, Eni and Black, Michael J.},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026}
}

MAMMA builds on a number of open-source models, datasets, and tools. We thank their authors for releasing their work openly.

• SAM 2 and SAM 3: Meta FAIR; segmentation backbones.
• YOLOv12: Ultralytics; person detection.
• SMPL-X: MPI-IS; expressive body model.
• ViTPose: pretrained human-pose backbone.
• HRNet: alternative pose-estimation backbone.
• CameraHMR: MPI-IS; landmark architecture baseline.
• BEDLAM: MPI-IS; synthetic dataset of humans in motion.
• Rerun: interactive 3D scene viewer.
• Detectron2: Meta FAIR; vision research library.
• PyTorch Lightning, Hydra, and WebDataset: training infrastructure.

For non-commercial scientific research purposes LICENSE.

Questions, bug reports, or other inquiries: mamma@tue.mpg.de.