Abstract
Whole-body SMPL-X estimators recover body, hand, and face meshes from single images, but applied frame by frame to video they produce jittery motion. Smoothing the output removes the jitter at a cost that body-only benchmarks do not measure: a temporal filter strong enough to steady the torso also flattens the fast finger strokes that carry meaning in sign language and gesture. We present CADENCE, a whole-body SMPL-X system for monocular video designed around this trade-off. A per-frame estimator built on a shared C-RADIOv4 encoder, with a high-resolution hand pathway that reuses the same encoder on hand crops, produces each frame's estimate. An offline temporal refiner then filters the SMPL-X rotation sequence: attention along the kinematic tree selects, for every joint and every frame, a mixture of learned zero-phase filters, so the torso is smoothed while fast finger strokes pass through. On 3DPW, the refiner reduces acceleration error from 17.4 to 5.4 mm/s² while lowering PA-MPJPE from 41.9 to 40.2 mm; generic smoothers applied to the same predictions reach 6.2 to 7.7 mm/s². On SignAvatars/How2Sign sign-language sequences, the refiner has the lowest hand jitter among the compared smoothers while keeping a hand-speed peak amplitude of 0.98, against 0.35 for a SmoothNet retrained on the full SMPL-X vector. The per-frame estimator also obtains the lowest full-body mesh error (25.7 mm) and hand joint error (8.3 mm) among the compared methods on EHF.
Smoothing without erasing the hands
Every smoother below takes the same per-frame CADENCE predictions on SignAvatars/How2Sign sign-language sequences. Moving left means less hand jitter; moving up means more of each hand stroke's peak speed survives (1.0 matches the reference). Generic smoothers buy stability by flattening strokes. A single-rate version of our refiner does the same. The per-joint bandwidth is what moves CADENCE up and to the left.
Hand jitter vs. hand-speed peak amplitude
- CADENCE refiner
- Other smoothers
- Per-frame input (no smoothing)
Data: Table 2 of the paper. Peak-A is measured against pseudo ground truth; the per-frame input overshoots it (1.18). Full numbers are in the table below.
Results
Temporal smoothing
| Method | PA-MPJPE ↓ | MPJPE ↓ | Accel ↓ |
|---|---|---|---|
| Per-frame | 41.9 | 61.7 | 17.4 |
| Gaussian (σ=2) | 41.5 | 61.1 | 7.0 |
| Moving average | 41.5 | 61.1 | 6.6 |
| One-Euro | 50.5 | 70.3 | 6.2 |
| Savitzky–Golay | 41.6 | 61.4 | 7.6 |
| SmoothNet-WB | 42.5 | 61.9 | 7.7 |
| Single-rate refiner | 41.7 | 61.2 | 5.7 |
| CADENCE refiner | 40.2 | 58.1 | 5.4 |
| Method | Jitter ↓ | Peak-R ↑ | Peak-A ↑ |
|---|---|---|---|
| Per-frame | 8.02 | 0.93 | 1.18 |
| SmoothNet-WB | 1.59 | 0.81 | 0.35 |
| Gaussian (σ=2) | 1.43 | 0.62 | 0.49 |
| One-Euro | 1.89 | 0.83 | 0.66 |
| Moving average | 2.22 | 0.78 | 0.58 |
| Savitzky–Golay | 3.25 | 0.83 | 0.76 |
| Single-rate refiner | 1.50 | 0.62 | 0.50 |
| CADENCE refiner | 1.29 | 0.90 | 0.98 |
Expressive whole-body accuracy (EHF)
| Method | P-MVE ↓ | P-MPJPE ↓ | |||
|---|---|---|---|---|---|
| Full | Hand | Face | Body | Hand | |
| PyMAF-X | 50.2 | 10.2 | 5.5 | 52.8 | 10.3 |
| SMPLer-X-L | 37.1 | 14.1 | 5.0 | – | – |
| Multi-HMR | 28.2 | 10.8 | 5.3 | – | – |
| NLF-L+fit | 26.0 | 20.7 | 6.3 | 26.1 | 21.8 |
| CADENCE | 25.7 | 12.2 | 5.1 | 20.6 | 8.3 |
On 3DPW, CADENCE's acceleration error (5.4) is lower than WHAM, TRAM, TCMR and MPS-Net (6.7–7.4) and slightly above the body-only GVHMR and GENMO (5.0, 5.2), while also predicting hands and face. Body-focused methods such as SAM 3D Body have lower body joint error. See the paper for the full comparison.
Qualitative results
BibTeX
@article{peart2026cadence,
title = {{CADENCE}: Temporally Coherent Whole-Body Human Mesh Recovery
that Preserves Expressive Hand Motion},
author = {Peart, Keniel and Chauhan, Jagmohan and Bodala, Indu P. and
Vishwakarma, Shelly},
journal = {arXiv preprint},
year = {2026}
}
Acknowledgements
The authors acknowledge the use of the IRIDIS X High Performance Computing Facility and the Southampton-Wolfson AI Research Machine (SWARM) GPU cluster, funded by the Wolfson Foundation, at the University of Southampton. SMPL-X, 3DPW, EHF, AGORA, UBody, and SignAvatars/How2Sign are used under their respective research licenses.