Simultaneous Person, Face, and Hand Detector Based on Improved YOLOv5

Many person, face, and hand detectors have reached remarkable achievements in the last century. With the convolutional neural networks' emergence, applications based on these detectors were increasingly widely applied in practice. Following this trend, this paper develops a simultaneous person, face, and hand detector based on YOLOv5 network architecture. The research focuses on redesigning the backbone and neck with compact network architectures EfficientNet and CBAM attention technique. The proposed architecture is trained and evaluated on the fine-tuned Human-Parts dataset. As a result, this network reached 87.2% of mAP@O.5 and 58.6% of mAP@O.5:0.95. This experimet outperforms most of the original YOLOv5 network architectures and is comparable to large-scale YOLOv5 architectures and the latest YOLOv8 architecture.