How to evaluate Object Detection models for your custom dataset

Scaled YOLOv4 vs YOLOv5

It is a very exciting time in the space of object detection models in computer vision as different research groups compete with each other to provide better approaches. The race over here is to generate models which can deliver better results in the tradeoff between accuracy, speed of inference, and the size of the model, The rationale being that object detection models need to be deployed in edge devices where both computation resources and system memory are in short supply, at the same time performance in terms of accuracy and speed is highly desired. Since the advent of YOLO in 2016, multiple generations of improvements such as YOLOv3, YOLOv4, YOLOv5, and Scaled YOLOv4 have been suggested by researchers. While there is a lot of work that has been done in evaluating and bench-marking the performance of these models, the reality is that results vary across datasets. Hence, if you are thinking about deploying an object detection model in production, you really need to evaluate how eac…