Machine-learning model annotation and training techniques

Abstract

A single item image is captured of an item situated within a given zone of a transaction area is captured, each different zone for each given item is associated a plurality of single item images captured by different cameras at different angles and perspectives of the transaction area. The single item images are passed to an existing segmentation Machine-Learning Model (MLM) and accurate masks for the items produced by the existing MLM are retained. A background image of an empty transaction area is obtained, each retained single item image is cropped and superimposed into the background image with one or more different cropped and superimposed singe item images creating a composite multi-item image. The composite multi-item images are labeled to identify the boundaries of each single item image and the existing segmentation MLM is trained on the composite multi-item and labeled images producing an enhanced segmentation MLM.