Pose-level synthetic data augmentation

One action.
A spectrum of variations.

Explore how real and generated human motions become new training examples.

Real repository dataInteractive 3D posesSaved results · no live inference

01 / From an existing example

Follow the motion.

See the pose extracted from one RGB video, then watch a human figure follow those same joints.

Original video

Source A

Extracted pose

RGB → 3D joints

Pose-driven human

Same extracted pose · fitted SOMA body

SOMA human Show joints on figurePose visualization · not augmentation

Sequence progressFrame 1 / 129

Uses source A of the example selected below. Both pose views show the same extracted motion; the variant slider does not change them. RGB follows normalized sequence progress, not verified event alignment.

02 / Explore the actual poses

Motion playground.

Motion example

Show Show pose jointsMotion preview Temporal smoothing for display · original data preserved

Frame 1 / 129Speed

Saved variant wA 0.5 · wB 0.5wA 0.1wA 0.9Camera angle 20°−180°Front 0°+180°Zoom 1×WideClose

The slider selects a saved optimizer output. Its parameters are not percentages of an endpoint crossfade.

About these motions and their provenance

The grid is a visual reference, not recovered ground geometry. Playback defaults to 25 fps for this preview. Joint coordinates retain the repository's root-relative representation.

SOMA bodies are fitted visualizations of the saved joints, with neutral hands and a fixed body shape. They are not recovered subject geometry. Original pose downloads are unchanged.

Loading verified motion examples…

03 / Motion studies and original renders