MAMMA tracks complex 3D motion with just 4 cameras

Hollywood mocap shoots have always looked ridiculous: tight bodysuits covered in ping-pong balls, surrounded by dozens of specialized infrared rigs. Even then, the moment two actors hugged or wrestled, the tracking points vanished and the data broke down.

Now, researchers from the Max Planck Institute for Intelligent Systems and Carnegie Mellon University have slashed that barrier. Their new project, MAMMA, manages accurate multi-person 3D motion capture using just four standard cameras.

Matrix bullet time with four iPhones

MAMMA stands for Markerless Accurate Multi-person Motion Acquisition, and the team is presenting it as a CVPR 2026 Oral paper.

Basically, it takes synchronized video from four angles—even outdoor footage shot on ordinary iPhones—and turns it into clean, expressive 3D human meshes. No wearable markers, no suits, and no studio tracking setups required.

Where it really shines is complex human interaction. Markerless systems usually choke when people touch, but MAMMA stays stable through close contact like dancing Bachata, hugging, and wrestling on the ground.

Because the full 3D geometry is reconstructed, the system can freely orbit around the subjects, letting you pull off seamless Matrix-style bullet time shots from minimal footage.

Turning frontier vision models into a unified pipeline

Instead of reinventing the wheel, the researchers cleverly orchestrated existing state-of-the-art vision models into a single modular pipeline: capture, masking, 2D landmarks, 3D optimization, and visualization.

Under the hood, it uses Meta FAIR's SAM 2 and SAM 3 for segmentation, YOLOv12 for detection, ViTPose and HRNet for joint extraction, and fits everything onto MPI's flagship SMPL-X parametric body model. 3D scenes are visualized natively with Rerun.

Even better, the release is remarkably polished for an academic paper. Alongside CLI runners, the GitHub repository includes a local web GUI built with Flask and React.

Fire up a dev script, navigate to port 3000 in your browser, and you can test a pre-packaged 56 MB four-camera sequence with one click. Five full dataset collections covering dance, multi-person interactions, and iPhone captures are also accessible on their project page.

Why it matters

We think MAMMA is a masterclass in modular engineering. It proves that combining specialized vision foundation models can outperform expensive, bulky hardware in real-world conditions.

For indie game devs, VFX animators, and digital artists, this drops the cost floor of multi-person mocap near zero. You do not need a multi-million-dollar Vicon stage to capture a two-person stunt; four mounted phones in an open park might soon be enough.

The catch right now is the license: it is strictly limited to non-commercial scientific research, meaning production studios cannot plug it straight into commercial pipelines yet. Multi-camera calibration and the 3D optimization pass also demand serious compute.

Still, the direction is clear. Marker-based capture suits are quickly becoming legacy technology, and general-purpose cameras powered by computer vision are taking over.