SMPL and SMPL-X Body Models

SMPL turns ten shape numbers plus joint angles into a posed 3D body mesh with 6,890 vertices.

Why Does This Exist?

Regressing $6{,}890 \times 3$ vertex coordinates freely invites impossible bodies. SMPL (2015) compresses humans into shape $\beta \in \mathbb{R}^{10}$ and pose $\theta \in \mathbb{R}^{72}$ for $24$ joints in axis-angle form. SMPL-X adds hand articulation and face expression on the same skeleton. Every human mesh recovery method outputs these numbers, not raw meshes.

Think of It Like This

A puppet with ten body sliders and joint strings

A puppet workshop keeps one template doll. Ten sliders widen hips or lengthen torsos, and strings set each of $24$ joints. Pulling strings bends the mesh by skinning without tearing. SMPL-X adds finger strings and face strings to the same doll.

Sliders are $\beta$ and strings are $\theta$ . The analogy stops at soft tissue: puppets do not jiggle while SMPL adds pose-dependent dents.

How It Actually Works

Template $\bar{T}$ plus shape blend $B_S(\beta)$ plus pose blend $B_P(\theta)$ gives unposed mesh $T$ . Linear blend skinning poses it: $v_i' = \sum_k w_{k,i} G_k(\theta) v_i$ with skinning weights $w$ and joint transforms $G_k$ . Joints come from linear regressor $J = \mathcal{J} T$ . SMPL has $6{,}890$ vertices; SMPL-X has $10{,}475$ with hands and face.

Worked example

Set $\beta = [1.0, 0, \ldots]$ widening hips $+3$ cm. Set elbow joint to $0.6$ rad flexion. A forearm vertex with weights $[0.8, 0.2]$ on elbow and wrist moves $0.8 \times$ elbow transform plus $0.2 \times$ wrist, landing $2.1$ cm forward. Reprojection of the wrist joint lands within $4$ pixels of detection. Zeroing $\beta$ returns the mean body exactly.

Watch Out For

Axis-angle jumps at 180 degrees

Pose near $\pi$ radians wraps discontinuously, so interpolation swings the long way. The symptom is arms corkscrewing between frames. Convert to rotation matrices for temporal smoothing.

The Quick Version

SMPL compresses bodies to $10$ shape plus $72$ pose numbers.
Skinning bends $6{,}890$ vertices from those numbers differentiably.
SMPL-X adds hands and face on one full-body rig.
Do not regress raw vertices when a prior exists; use the parameters.
Ten shape dials cover most adult variation for fitting.