Customized neural network for item recognition
A camera acquires an image of an item involved in an interaction, such as the item being picked or placed on a shelf by a user. The item depicted in the image is identified via item recognition using a customized neural network that uses previously trained subnetworks for a set of candidate items. The candidate items are determined based on items near a user location, items associated with a user cart, items in a shopping list or wish list associated with the user, or items otherwise associated with the user. Each previously trained subnetwork for a candidate item includes one or more reference images and corresponding reference feature data. The customized neural network uses a matching and inference network to test the image against the candidate items and provide comparison data indicative of a similarity between the item depicted in the image and one of the candidate items.
1. A system comprising:
a shelf;
a camera having a field-of-view that includes the shelf; and
a computing device comprising:
a memory storing first computer-executable instructions; and
a hardware processor to execute the first computer-executable instructions to:
determine a location of a cart is within a first threshold distance of a first location of a user;
determine the cart is associated with the user based on the location of the cart being within the first threshold distance of the first location of the user;
access a first image acquired by the camera;
determine a second location of the user is within a second threshold distance of a location of the shelf;
determine cart data, the cart data comprising data indicative of items stowed in the cart;
determine candidate item data based on the second location of the user being within the second threshold distance of the location of the shelf and based on the cart being associated with the user, the candidate item data comprising data indicative of one or more items, wherein the one or more items include the items stowed in the cart;
instantiate a neural network comprising:
a previously trained subnetwork for each item of the one or more items, wherein the previously trained subnetwork for the each item comprises:
a reference image for the each item;
a feature network corresponding to the reference image for the each item; and
a matching network for the each item;
a test feature network; and
an inference network;
generate, by operating the neural network, test feature data associated with the first image;
determine, by operating the neural network, reference feature data for the reference image for the each item;
generate, by operating the neural network, matching data based on the test feature data and the reference feature data for the reference image for the each item; and
determine, using the matching data for the each item, comparison data indicative of a correspondence between the first image and one or more of the one or more items.
2. The system of claim 1 , wherein the one or more items further include one or more of:
first items included in a shopping list associated with the user, or
second items included in a wish list associated with the user.
3. The system of claim 1 , wherein the one or more items further include items that the user previously interacted with before the first image was acquired.
4. The system of claim 1 , the hardware processor to further execute the first computer-executable instructions to:
determine account data associated with the user;
determine a prediction of items to be purchased by the user, based on the account data; and
wherein the one or more items further include the items predicted to be purchased by the user.
5. The system of claim 1 , further comprising:
the memory storing second computer-executable instructions; and
the hardware processor to execute the second computer-executable instructions to:
access training data comprising training images of the each item in different poses;
instantiate a second neural network;
train the second neural network to recognize a match between an item depicted in a test image and the each item depicted in the training images; and
store parameter data comprising parameters of the feature network and the matching network for the previously trained subnetwork for the each item.
6. The system of claim 1 , further comprising:
the memory storing second computer-executable instructions; and
the hardware processor to execute the second computer-executable instructions to:
instantiate a multi-layer perceptron for the each item, wherein every instantiation of the multi-layer perceptron uses a same set of one or more predetermined parameters; and
process, with a respective multi-layer perceptron, the reference feature data of the each item to generate the matching data for the each item.
7. A method comprising:
determining a location of a cart is within a first threshold distance of a first location of a user;
determining the cart is associated with the user based on the location of the cart being within the first threshold distance of the first location of the user;
accessing a first image;
determining the first image is associated with an inventory location;
determining a second location of the user is within a second threshold distance of the inventory location;
determining cart data, the cart data comprising data indicative of items stowed in the cart;
determining candidate item data based on the second location of the user being within the second threshold distance of the inventory location and based on the cart being associated with the user, the candidate item data comprising data indicative of one or more items associated with the user, wherein the one or more items associated with the user include the items stowed in the cart;
instantiating a neural network comprising:
a previously trained subnetwork for each item of the one or more items, wherein the previously trained subnetwork for the each item comprises:
a reference image for the each item;
a feature network corresponding to the reference image for the each item; and
a matching network for the each item;
a test feature network; and
an inference network;
generating, using the neural network, test feature data associated with the first image;
generating, using the neural network, matching data based on the test feature data and reference feature data associated with the each item; and
determining, using the matching data for the each item, comparison data indicative of a correspondence between the first image and at least one of the one or more items.
8. The method of claim 7 , further comprising:
determining a shopping list of items associated with the user; and
designating the items in the shopping list as candidate items.
9. The method of claim 7 , further comprising:
determining items previously purchased by the user; and
designating the items previously purchased by the user as candidate items.
10. The method of claim 7 , further comprising:
accessing account data associated with the user;
determining a prediction of items to be purchased by the user, based on the account data; and
designating the items predicted to be purchased by the user as candidate items.
11. The method of claim 7 , further comprising:
determining items designated for stowage at the inventory location; and
designating the items designated for stowage at the inventory location as candidate items.
12. A system comprising:
a computing device comprising:
a memory storing first computer-executable instructions; and
a hardware processor to execute the first computer-executable instructions to:
determine a location of a cart is within a first threshold distance of a first location of a user;
determine the cart is associated with the user based on the location of the cart being within the first threshold distance of the first location of the user;
access a first image acquired by a camera;
determine the first image is associated with an inventory location;
determine a second location of the user is within a second threshold distance of the inventory location;
determine cart data indicative of items stowed in the cart;
determine candidate item data based on the second location of the user being within the second threshold distance of the inventory location and based on the cart being associated with the user, the candidate item data comprising data indicative of one or more items associated with the user, wherein the one or more items include the items stowed in the cart;
instantiate a neural network comprising:
a previously trained subnetwork for each item of the one or more items, wherein the previously trained subnetwork for the each item comprises:
a reference image for the each item;
a feature network corresponding to the reference image for the each item; and
a matching network for the each item;
a test feature network; and
an inference network;
determine, by operating the neural network, test feature data associated with the first image;
determine, by operating the neural network, reference feature data for the reference image for the each item;
generate, by operating the neural network, matching data based on the test feature data and the reference feature data for the reference image for the each item; and
determine, using the matching data for the each item, comparison data indicative of a correspondence between the first image and one or more of the one or more items.
13. The system of claim 12 , the hardware processor to further execute the first computer-executable instructions to:
determine that an item depicted in the first image matches an item depicted in one of the reference images, based on the comparison data.
14. The system of claim 12 , the hardware processor to further execute the first computer-executable instructions to:
access account data associated with the user;
determine a prediction of items to be purchased by the user, based on the account data; and
wherein the one or more items further include the items predicted to be purchased by the user.
15. The system of claim 12 , wherein the previously trained subnetwork for the each item comprises a convolutional neural network with a residual network architecture in which one or more predetermined parameters are identical for each instantiation.