CPC G06N 3/08 (2013.01) [G06Q 30/0631 (2013.01); G06Q 30/0643 (2013.01)] | 15 Claims |
1. A computer-implemented method comprising:
obtaining a pair of images of apparel items;
providing the pair of images as input to a trained neural network, wherein the trained neural network includes a first subnetwork comprising a left branch and a right branch;
processing the pair of images with the trained neural network, wherein processing the pair of images includes:
generating first feature embeddings for a first image of the pair of images using the left branch of the first subnetwork and generating second feature embeddings for a second image of the pair of images using the right branch of the first subnetwork;
using a combiner, generating a combined feature vector of the first feature embeddings and the second feature embeddings by calculating a Hadamard product of the first feature embeddings and the second feature embeddings; and
mapping the combined feature vector to a compatibility score using a readout function of a second subnetwork, the second subnetwork including a fully connected (FC) neural network;
obtaining the compatibility score as output from the second subnetwork of the trained neural network; and
upon determining that the compatibility score is over a predetermined threshold value, recommending that a first apparel item associated with the first image and a second apparel item associated with second image are compatible.
|