Gun Detection: Comparative Analysis using Transfer Learning in Single Stage Detectors

Abstract

Every year, a lot of people around the world suffer from gun-related violence. A solution for this could be using a Single stage detector to detect such incidents quickly. They provide accurate and fast detection. Normally in single stage detectors YOLOv3tiny provides fast detection than YOLOv3 but with less accuracy. But in this paper when transfer learning is applied to both the versions with the small dataset having new class as gun then tiny version improves with accuracy by 4% than that of v3. When YOLOv3 and tiny version are trained on 3000 and 2500 respectively then we have got that point as a threshold where both gave same accuracy. Their performances were also evaluated using criteria such as precision, recall, F1 score. The key takeaway from this is YOLOv3 tiny performed best in terms of accuracy and F1 score than that of YOLOv3 in case of transfer learning.