cuevhv/mamma

Official code for MAMMA: Markerless Accurate Multi-person Motion Acquisition.

What it solves

MAMMA 提供了一种方法，能够从多视角视频中准确捕获多人的 3D 动作，且不需要在受试者身体上使用实体标记。

How it works

The system uses a multi-stage pipeline to process video captures:

  1. Capture (ma_cap): Loads the multi-view video data.
  2. Segmentation (ma_masks): Uses SAM and YOLO to isolate individual people in the frames.
  3. 2D Landmark Detection (ma_2d): Employs a specialized network called MammaNet to detect 2D body landmarks.
  4. 3D Optimization (ma_3d): Performs multi-view SMPL-X optimization to reconstruct the 3D body poses.
  5. Visualization (ma_vis): Generates per-camera overlays and an interactive 3D scene.

Who it’s for

Researchers and developers in computer vision, animation, and motion capture who need high-fidelity 3D human motion data from standard multi-camera setups.

Highlights

  • Markerless approach: Eliminates the need for specialized markers on subjects.
  • Multi-person support: Capable of tracking and reconstructing multiple people simultaneously.
  • Integrated GUI: Includes a browser-based interface for submitting and inspecting runs.
  • Comprehensive Pipeline: Combines segmentation, 2D pose estimation, and 3D optimization into a single workflow.