A large, crowdsourced evaluation of gesture generation systems on common data: The GENEA ...
summary
A community benchmark generates gestures from a common ~4-hour speech+3D-motion dataset, renders every entry on the same avatar, and evaluates them with a crowdsourced interface; HEMVIP interface collects 125 attentive raters per study. Baselines use a denoising autoencoder + Savitzky‑Golay smoothing (audio → motion) and a text‑encoder with FastText; 15 registrations yield 5 submissions and 2 open repos. Natural motion still outperforms all systems, but mismatched motion rated highly (second best median), and common objective metrics fail to predict perceptual quality.
AI-written summary · about
related talks
19mA Rapid Prototyping Approach to Synthetic Data Generation For Improved 2D Gesture Recognition
06mGAN-based Style Transformation to Improve Gesture-recognition Accuracy
19mChallenges in Generated Video Evaluation
06mPantomime: Mid-Air Gesture Recognition with Sparse Millimeter-Wave Radar Point Clouds
05mA Feature Adaptive Learning Method for High-Density sEMG-Based Gesture Recognition
09mBeyond Captions A Practical Pipeline for Live AI-Generated Sign Language Avatars
22mStrike a Pose: Gesture Recognition in JS with Machine Learning
15mRecognizing Emotional Expression in Game Streams