These 18 methods cover every hand pose source in the 22 Taxel Atlas datasets, plus MediaPipe and HaWoR, which no catalog dataset uses yet but which are common ways to add pose to video. They are sorted on two axes: the hardware a recording needs, from any video to a worn glove or a robot, and how the pose is computed from the raw signal: a learned model, multi-view geometry, inertial fusion, a magnetic field, or a calibrated reading per joint.
Each entry draws what goes in and what comes out, with array shapes and units, and lists the published accuracy figures with the benchmark or reference system behind them.
Rows are the hardware a recording needs, columns the way the pose is computed. A dashed name is a method whose second stage uses that column too. Each name links to its entry below; datasets with no hand pose are listed at the end.
| Method | Computation | Output | Frame | Published error | Datasets in the catalog |
|---|
Each figure below comes from its own reference and conditions, so numbers compare within one method's list, and only loosely across methods.
Teleoperated and human-to-robot datasets turn a tracked human hand into robot joint angles. The released robot joints are measured, but how well they match what the operator did depends on this mapping. The most common open tool is dex-retargeting from AnyTeleop (Qin et al., RSS 2023): it solves for joint angles that match human keypoints, in a vector mode suggested for teleoperation, a position mode suggested for offline human data, or a dexpilot mode that follows the DexPilot cost.
| Dataset | Human input | Mapping | Robot |
|---|
Method facts and error figures were checked on 2026-10-16 against the papers, vendor specifications and SDK documentation linked under each entry. Vendor figures are marked as vendor claims. Dataset details come from each dataset's catalog entry. Where a paper and a vendor disagree (TactiDex and DexViTac call MANUS EMF gloves "IMU" gloves), both are stated.