Online Hand Gesture Recognition Using 3D CNNs
A new system for real-time hand gesture recognition using 3D convolutional neural networks achieves over 98% accuracy on the Jester dataset for detection and over 90% for classification. The system employs a sliding window approach to refine results from multiple windows, enabling localization and recognition of gestures in live video streams. On a homemade dataset, the best configuration reaches 37.5% Levenshtein accuracy with a response time under three seconds. The project aims to address challenges in human-computer interaction, such as real-time performance and variability in gesture execution. The code is publicly available.
Key facts
- System uses 3D convolutional neural networks for hand gesture recognition.
- Achieves 98+% accuracy for detector and 90+% for classifier on Jester database.
- Sliding window approach refines results from multiple windows.
- Best group responds within three seconds on homemade dataset.
- 37.5% Levenshtein accuracy on homemade dataset.
- Designed for real-time video stream processing.
- Addresses variability in how people perform gestures.
- Project codes are publicly available.
Entities
Institutions
- arXiv