Authors:
Srija Ganguly, Evgeny Shevtsov, Andrei Khaitas
Abstract:
The point-of-sale system has just been introduced with Produce Recognition features. Even after a produce is recognized, the shopper has to sometimes input the quantity, or they must wait for the system to weigh the item after it has been added to the transaction.
This solution aims to provide a method for the system to predict the weight or quantity for the items on its own without the shopper having to perform any extra steps.
Background:
At the self-checkout terminal, often an item needs to be weighed, or the quantity needs to be input for the system to add that item to the transaction. Nowadays one of the major retail stores have started using a produce recognition system (which is extensible to non-produce items too depending on the enrollment data) where an item would need to be placed on the scanner scale and the shopper gets predictions on the screen for what the camera might see, but even so they have to either wait for the system to weigh the item or enter the quantity manually in a separate step. This additional step could be eliminated if the recognition system could be trained to suggest a weight or quantity depending on the item that is being held in front of the camera. This would likely eliminate the requirement of a scale at the point-of-sale system too.
Description:
Nowadays one of the major retail stores have started using a produce recognition system (which is extensible to non-produce as well) where an item would need to be placed on the scanner scale and the shopper gets predictions on the screen for what the camera might see, but even so they have to either wait for the system to weigh the item or enter the quantity manually in a separate step. This additional step could be eliminated if the recognition system could be trained to suggest a weight or quantity depending on the item that is being held in front of the camera. There would be some mathematical calculations that would be used by the neural network model to precisely identify a possible weight or quantity depending on the mass that the recognition system sees. Of course, the algorithms for identifying the quantity would be another recognition feature that would have to predict the number based on the item's weight and related training data (for example) and the weight can be predicted by using some math, grid-based mapping etc. This would likely eliminate the requirement of a scale at the point-of-sale system too.
The intended flow for a weight enabled product would be -
Shopper wakes up the Self-Checkout Application -- they place a banana on a normal tray (that need not be a scale) -- camera captures the item, and the recognition model sends predictions to the screen. It also internally maps the possible weight of the banana using training data, grids, visible mass or any other relevant algorithm as implemented -- shopper clicks on the most relevant prediction of Bananas -- in the next step, instead of the system trying to weigh the banana, the recognition model directly adds the selected prediction with its pre-identified weight to the transaction.
The intended flow for a quantity enabled product would be -
Shopper wakes up the Self-Checkout Application -- they place a pepper on a normal tray (that need not be a scale) -- camera captures the item, and the recognition model sends predictions to the screen. It also internally maps the possible number of peppers using weight training data, grids, visible mass or any other relevant algorithm as implemented -- shopper clicks on the most relevant prediction of Peppers -- in the next step, instead of the system asking them to enter the quantity of the peppers, the recognition model directly adds the selected prediction with its predicted quantity to the transaction.
This would not only reduce the number of steps a shopper has to take to complete their transaction but also helps in reducing certain hardware and system integrations which may be required to accommodate the weighing step.
TGCS Reference 3025
