Machine-learning model annotation and training techniques

A method, comprising:

• providing single item images as input to an existing segmentation Machine-Learning Model (MLM), wherein each single item image depicts a single item within a transaction area;
• receiving a mask as output from the existing segmentation MLM for each single item image, wherein each mask identifies a set of pixels that represents a corresponding single item within a corresponding single item image;
• obtaining a background image depicting an empty version of the transaction area without any item being depicted;
• generating composite multi-item images by using a corresponding mask for a corresponding single item and cropping each corresponding set of pixels from the corresponding single item image creating a cropped item image and placing the cropped item image in combinations within one or more other cropped item images into the background image, wherein each composite multi-item image depicts two or more single items within the transaction area;
• labeling boundaries for each cropped item image within a corresponding composite multi-item image based on a corresponding mask and a location that a corresponding cropped item image was placed in the corresponding composite multi-item image and creating a labeled multi-item scene image for each composite multi-item image;
• and
• deriving an enhanced segmentation MLM by training the existing segmentation MLM on each labeled multi-item scene image;
• wherein the enhanced segmentation MLM is trained on multiple different combinations of items, each combination varies in both a number of items and actual items used during each iteration.