Artificial intelligence powered styling agent
A system comprising one or more processors and one or more non-transitory computer-readable media storing computing instructions, that when executed on the one or more processors, cause the one or more processors to perform: receiving stock images comprising an anchor garment; automatically identifying the anchor garment and complementary garments within the stock images; selecting an image of the stock images in which a mask area of a first complementary garment of the complementary garments as a ratio of an area of the anchor garment is largest over other complementary garments of the complementary garments; performing an image search, using the image, in an item catalog for similar garments to the first complementary garment; and displaying, on a user interface, an avatar wearing the anchor garment and at least one of the similar garments. Other embodiments are disclosed.
1 . A system comprising:
one or more processors; and
one or more non-transitory computer-readable media storing computing instructions, that when executed on the one or more processors, cause the one or more processors to perform:
receiving stock images comprising an anchor garment;
automatically identifying the anchor garment and complementary garments within the stock images;
selecting an image of the stock images in which a mask area of a first complementary garment of the complementary garments as a ratio of an area of the anchor garment is largest over other complementary garments of the complementary garments, wherein the mask area comprises a pixel-level segmentation mask area generated by a semantic or instance segmentation model;
performing an image search, using the image, in an item catalog for similar garments to the first complementary garment; and
displaying, on a user interface, an avatar wearing the anchor garment and at least one of the similar garments.
2 . The system of claim 1 , wherein automatically identifying the anchor garment and the complementary garments comprises:
using a segmentation model to identify the anchor garment and the complementary garments within the stock images.
3 . The system of claim 2 , wherein automatically identifying the anchor garment and the complementary garments comprises:
identifying the anchor garment based on which garment is most commonly found in the stock images.
4 . The system of claim 1 , wherein selecting the image comprises:
filtering out the stock images in which the complementary garments are partially cropped out.
5 . The system of claim 1 , wherein performing the image search further comprises:
pre-training a visual search model;
performing deep clustering on the visual search model, as pre-trained, to mine k-nearest neighbors, with hard negative mining based on garment metadata; and
performing active learning.
6 . The system of claim 5 , wherein pre-training the visual search model further comprises:
augmenting batch images for training the visual search model with positive examples or negative examples.
7 . The system of claim 6 , wherein augmenting the batch images comprises:
generating new images to be the positive examples, based on the stock images that comprise the first complementary garment, by at least one of:
changing hues of the first complementary garment;
changing an angle of or skewing the first complementary garment;
changing a size of the first complementary garment;
adding holes in the stock images of the first complementary garment; or
changing an avatar model wearing the first complementary garment using a virtual try on (VTO) model.
8 . The system of claim 6 , wherein augmenting the batch images further comprises:
automatically selecting the negative examples from images of other garments in the item catalog.
9 . The system of claim 6 , wherein augmenting the batch images further comprises:
generating new images to be the negative examples, based on the stock images that comprise the first complementary garment, by changing a color of the first complementary garment.
10 . The system of claim 5 , wherein performing the active learning comprises:
submitting style proposals to individuals for feedback, wherein the style proposals each comprise the anchor garment and at least one of the similar garments as a group;
receiving feedback from the individuals; and
using the style proposals that are rejected as negative examples in a feedback loop.
11 . A method being implemented via execution of computing instructions configured to run on one or more processors and stored at one or more non-transitory media, the method comprising:
receiving stock images comprising an anchor garment;
automatically identifying the anchor garment and complementary garments within the stock images;
selecting an image of the stock images in which a mask area of a first complementary garment of the complementary garments as a ratio of an area of the anchor garment is largest over other complementary garments of the complementary garments, wherein the mask area comprises a pixel-level segmentation mask area generated by a semantic or instance segmentation model;
performing an image search, using the image, in an item catalog for similar garments to the first complementary garment; and
displaying, on a user interface, an avatar wearing the anchor garment and at least one of the similar garments.
12 . The method of claim 11 , wherein automatically identifying the anchor garment and the complementary garments comprises:
using a segmentation model to identify the anchor garment and the complementary garments within the stock images.
13 . The method of claim 12 , wherein automatically identifying the anchor garment and the complementary garments comprises:
identifying the anchor garment based on which garment is most commonly found in the stock images.
14 . The method of claim 11 , wherein selecting the image comprises:
filtering out the stock images in which the complementary garments are partially cropped out.
15 . The method of claim 11 , wherein performing the image search further comprises:
pre-training a visual search model;
performing deep clustering on the visual search model, as pre-trained, to mine k-nearest neighbors, with hard negative mining based on garment metadata; and
performing active learning.
16 . The method of claim 15 , wherein pre-training the visual search model further comprises:
augmenting batch images for training the visual search model with positive examples or negative examples.
17 . The method of claim 16 , wherein augmenting the batch images comprises:
generating new images to be the positive examples, based on the stock images that comprise the first complementary garment, by at least one of:
changing hues of the first complementary garment;
changing an angle of or skewing the first complementary garment;
changing a size of the first complementary garment;
adding holes in the stock images of the first complementary garment; or
changing an avatar model wearing the first complementary garment using a virtual try on (VTO) model.
18 . The method of claim 16 , wherein augmenting the batch images further comprises:
automatically selecting the negative examples from images of other garments in the item catalog.
19 . The method of claim 16 , wherein augmenting the batch images further comprises:
generating new images to be the negative examples, based on the stock images that comprise the first complementary garment, by changing a color of the first complementary garment.
20 . The method of claim 15 , wherein performing the active learning comprises:
submitting style proposals to individuals for feedback, wherein the style proposals each comprise the anchor garment and at least one of the similar garments as a group;
receiving feedback from the individuals; and
using the style proposals that are rejected as negative examples in a feedback loop.