@tejasmhos check this out. The COCO model detected boxes well but failed on bags. Crazy how it recognized the pizza and bowl on the labels. I'll work on batch inference and dissect the predicted class labels (i think right now it only outputs top class prediction)

That's amazing! One option to try out is outputting box whenever it is among the predicted classes, even if it is not the top prediction. I'm working on collecting images for the synthetic generation step for boxes to begin with.