Back to Academic/Research
Research PreviewSpecializedvision3dvideoUpdated July 15, 2026

MAMMA

Model Overview

MAMMA is a undisclosed-parameter model developed by Academic/Research. Released on 2026-07-15.


📊 Quick Specs

Specification Table
SpecificationValue
Parametersundisclosed
Taskother
Modalityvision, 3d, video
LicenseOther/Custom
Typeresearch-preview

✨ Key Features

  • MammaNet Transformer Architecture: Employs a ViT-Base backbone to decode 512 dense surface landmarks with pixel coordinates, uncertainty, and visibility estimation
  • Contact-Aware & Visibility Heads: Jointly predicts person-to-person and person-to-floor contact probabilities to constrain 3D multi-person pose optimization
  • Markerless SMPL-X Fitting: Recovers full-body expressive SMPL-X parameters directly from synchronized multi-view video without specialized marker suits
  • Multi-View Epipolar Optimization: Uses symmetric epipolar distance matching across camera views to resolve person-to-landmark correspondences under heavy occlusion
  • MammaSyn Synthetic Dataset: Trained on a large-scale synthetic multi-person dataset with ground-truth dense landmarks and physics-based contact labels
  • Web GUI & CLI Tooling: Offers zero-config command-line execution (python -m inference run) as well as a local web browser user interface on port 3000

🔗 Resources


📜 License & Access

Other/Custom — See repository for specific license details.

Key Features

MammaNet Transformer Architecture: Employs a ViT-Base backbone to decode 512 dense surface landmarks with pixel coordinates, uncertainty, and visibility estimation

Feature 01

Contact-Aware & Visibility Heads: Jointly predicts person-to-person and person-to-floor contact probabilities to constrain 3D multi-person pose optimization

Feature 02

Markerless SMPL-X Fitting: Recovers full-body expressive SMPL-X parameters directly from synchronized multi-view video without specialized marker suits

Feature 03

Multi-View Epipolar Optimization: Uses symmetric epipolar distance matching across camera views to resolve person-to-landmark correspondences under heavy occlusion

Feature 04

MammaSyn Synthetic Dataset: Trained on a large-scale synthetic multi-person dataset with ground-truth dense landmarks and physics-based contact labels

Feature 05

Web GUI & CLI Tooling: Offers zero-config command-line execution (python -m inference run) as well as a local web browser user interface on port 3000

Feature 06

You might also want to compare

Verified Sources

Tags

research-previewmotion-capturesmpl-xcomputer-vision3d-pose-estimationmulti-person-interactioncvpr-2026

Model Specs

research-preview

Parameters

Undisclosed

Context Window

undisclosed

License

Other/Custom

Deployment

self-hostable

Resources & Links

Curator Notes

Partially enriched via migration on 2026-07-25. Manual review recommended.

Compare Specs

Compare parameters, context windows, modalities, and benchmark scores of this model side-by-side with others.

Compare Model