Jester-Informed Synthetic Data Generation for TPU-Accelerated Binary Gesture Recognition

This work presents a binary gesture recognition system that overcomes the usual real-world data requirements by training only on synthetic videos based on the Jester dataset. Instead of relying on large annotated video datasets, the motion patterns of the 20BN-Jester dataset are analyzed to produce parametrically controlled swipe-left gesture sequences based on real-world human motion statistics. A specially designed 3D convolutional neural network with 2.16 million parameters is trained on the synthetic data using Google TPU v3-8 hardware. The synthetic data generation process simulates hand trajectories, temporal patterns, and real-world movement variations as seen in real Jester videos, allowing for a strong correspondence between synthetic and real-world gesture behaviors. On synthetic validation tasks, the network converges in the first training iteration and reaches full validation accuracy. On a held-out set of 23 real Jester swipe-left videos, the same network demonstrates strong synthetic-to-real transfer with excellent generalization performance. The proposed system overcomes privacy concerns in data collection, supports balanced training data, and improves computational efficiency compared to traditional real-data systems. A lightweight supplementary implementation further verifies strong convergence performance in resource-constrained hardware settings.