Technically this is a pill bottle detector which was derived from a custom object detector trained by yolov5. Identifying the pill bottles dedicated for family members is challenging, particularly when people are getting old. E.g, they often forget where the bottle is, or they cannot ensure if the pill bottle is dedicated for someone when the bottles were acquired from the same pharmacy although they were labeled differently. Applying one of the state-of-the-art object detectors i.e yolov5 to train a custom object detector might help people address this issue.
Therefore, this code snippet is aimed at training and testing a custom object detection process which distinguishes different pill bottles even though they sometimes look alike.
What made it different is that this code leverages the OpenCV DNN backend as the Deep Learning backbone. It works in a way that is different than a regular YOLOv5 detector, which typically uses the detect.py developed by Ultralytics.
In order to do so, an onnx-format weights file was converted from the pytorch's dot-pt-format which was taken by the transfer learning proccess. From this standpoint, it is an indepentent app for detecting pill bottles.
Another important step that this app has made is that it developed a pipeline which has figured out the whole development processs from capturing training photos, labeling and resizeing the training data which comply with the requirement of Yolov5. The labeling tool that has been used in this case particularly is not the official recommended Roboflow annotator, instead, The Label Studio which is an open-sourced application has been used to label the dataset while still protecting the personal training data from leaking to public, which is the typical drawback of Roboflow.
