Realtime Object Detection of Pantry Objects Using YOLOv5 Transfer Learning in Varying Lighting and Orientation

Abstract

This paper describes the use of YOLOv5 transfer learning from the COCO dataset to train and deploy a custom model to detect select pantry objects in various lighting and orientations using an original custom dataset with applied brightness and saturation augmentations. The results show that the trained model using the custom dataset obtained an mAP(0.5) of 0.9948 at 87 epochs of training and an accuracy of 100% detections and 95% average confidence level for the Koko Krunch set and an accuracy of 100% detections and 91% average confidence level for the Lady's Choice Mayonnaise set during validation.