A Comparative Study of YOLOv5 models on American Sign Language Dataset

Sign language is the most common way of communication for people with hearing and speech difficulties. One of the biggest problems for sign language user is that most people does not understand sign language. The most promising solution to this problem is a sign language detection system using object detection algorithm. YOLOv5 is state-of-Art one-stage object detection algorithm and is available in wide range of model complexity, ranging from simplest YOLOv5n to most complex YOLOv5x. To achieve efficient communication, the sign language detection system needs to be fast and reliable. However, many previous studies only used the most complex model without considering the time needed for the system to run. As more complex model tends to performs better at the cost of computational time, the most optimal model for sign language detection system is a model that performs well while maintaining fast inference time. In this study, we compare the inference time and performance of every YOLOv5 model available, trained on American Sign Language dataset to find the most optimal model of YOLOv5 for sign language detection. The experiment results shows that while YOLOv5x has slightly better performance than other models with mAP of 0.88 and F1 score of 0.91, it required twice the amount of time to detect the sign language with inference time of 26.2 ms. The same can be said to YOLOv5m and YOLOv5l, both with mAP of 0.88 and F1 score of 0.88 and 0.90, while require inference time of 16.2 ms and 19.1 ms respectively. YOLOv5n is the fastest model at inference time of 7.2 ms, but the performance is considerably worse with mAP of 0.79 and F1 score of 0.88. In conclusion, YOLOv5s is the most optimal model with mAP of 0.88, F1 score of 0.90, and inference time of 10.6 ms.