A Deployment Scheme of YOLOv5 with Inference Optimizations Based on the Triton Inference Server

Object detection constitutes a large part of computer vision applications. You Only Look Once (YOLO) v5 is a salient object detection algorithm that provides high accuracy and real-time performance. This paper illustrates a deployment scheme of YOLOv5 with inference optimizations on Nvidia graphics cards using an open-source deep-learning deployment framework named Triton Inference Server. Moreover, we developed a non-maximum suppression (NMS) operator with dynamic-batch-size support in TensorRT to accelerate inference. The experimental results show that both throughput and latency are improved significantly through our deployment scheme.