GAN-based Style Transformation to Improve Gesture-recognition Accuracy
summary
A GAN-based style transformer synthesizes missing personal gesture samples by performing intra- and inter-user transformations and trains personal classifiers with the generated examples. The system uses five losses (adversarial, style class, content class, cycle-consistency, smoothness) and is tested with SVM and CNN on CheekInput and USC-HAD. CNN accuracy rises from 0.747 to 0.822 on CheekInput and from 0.856 to 0.899 on USC-HAD, achieving similar performance using only 25% and 43% of target-user data respectively.
AI-written summary · about
related talks
10mDeepNAG: Deep Non-Adversarial Gesture Generation
13m[Interactive] Neural Style Transfer: The Art and Science of Generative AI
58mStyle Transfer AI - How Neural Networks Generate Art
19mA Rapid Prototyping Approach to Synthetic Data Generation For Improved 2D Gesture Recognition
03mBuilding artificial people: Endless possibilities and the dark side
15mDrag Makeovers, via AI!
37mHarnessing the power of Generative Adversarial Networks (GANS) for supervised learning
10mA large, crowdsourced evaluation of gesture generation systems on common data: The GENEA ...