Enhancing Gesture Recognition Performance Using Optimized Event-Based Data Sample Lengths and Crops

In recent years, event-based cameras have found their use in a growing number of real-time and mobile computer vision applications. This technology has the potential of improving sign language gesture recognition, a complex visual recognition task that benefits from capturing high-speed motion. Focusing on three spiking neural network (SNN) recognition systems – STBP, SLAYER and DECOLLE – applied on the SL-Animals-DVS dataset, this study investigates novel approaches for preprocessing Dynamic Vision Sensor input signals and modifying network architectures to enhance training speed, reduce memory load, and improve recognition performance. The experiments reveal that optimizing sample lengths and employing random crop techniques during training significantly enhance SNN performance. While optimizing computational resources, this work emphasizes the importance of exposing SNNs to more information from training samples, highlighting the trade-off between sample length, performance, memory usage, and training time.