ahmetgunduz/Real-time-GesRec

Implements a two-stage hierarchical architecture combining lightweight detection with deep 3D CNN classifiers (ResNet-10, ResNeXt-101, C3D) to enable online gesture recognition via sliding windows, addressing temporal boundary detection without explicit frame annotations. Supports both RGB and depth modalities across multiple datasets and introduces Levenshtein distance as an evaluation metric to capture misclassifications, duplicates, and missed detections in continuous video streams. Includes pretrained models and utilities for dataset preparation on EgoGesture, nvGesture, Jester, Kinetics, and UCF101.

691 stars. No commits in the last 6 months.