Visual attribute expansion via multiple machine learning models
A computer implemented method includes receiving an image that includes a type of object, segmenting the object into multiple segments via a trained segmentation machine learning model, and inputting the segments into multiple different attribute extraction models to extract different types of attributes from each of the multiple segments.
1 . A computer implemented method comprising:
receiving an image that includes a type of object;
segmenting the image of the object into multiple segments, including multiple body part shape segments, via a trained segmentation machine learning model; and
inputting the segments into multiple different attribute extraction models to extract different types of attributes from each of the multiple segments.
2 . The method of claim 1 wherein the trained segmentation machine learning model has been trained on multiple images of the type of object that are labeled with multiple tagged segments identified by bounding boxes.
3 . The method of claim 2 wherein the type of object comprises a bottle and wherein tags comprise classes including bottle, neck, shoulder, body, top, and logo.
4 . The method of claim 3 wherein the top class includes multiple different top classes and wherein the logo class includes multiple different logo classes.
5 . The method of claim 1 wherein the trained segmentation machine learning model comprises a mask-recurrent convolutional neural network having classes corresponding to the multiple segments of the object.
6 . The method of claim 1 wherein the attribute extraction models include a shape attribute extraction model and a color attribute extraction model.
7 . The method of claim 6 wherein the color attribute extraction model generates image embedding on each salient region of the image and determines distance differences between the image embeddings and target color text embeddings in the same embedding space, to output target colors.
8 . The method of claim 7 wherein the color attribute extraction model is an unsupervised model and wherein a distance difference is compared to a threshold to determine that a color is present in a salient region corresponding to one of the segments.
9 . The method of claim 7 wherein the color attribute extraction model generates a list of colors in descending order of area covered by each color in a salient region.
10 . The method of claim 6 wherein the shape attribute model utilizes ratios of measurements of bounding boxes corresponding to segments to generate shape descriptions.
11 . The method of claim 6 wherein the attribute extraction models include a design elements attribute extraction model.
12 . The method of claim 11 wherein the design elements attribute extraction model is an unsupervised model that compares image embeddings to design element text embeddings in the same embedding space to select design element themes.
13 . The method of claim 12 wherein the type of object comprises a bottle and wherein the design elements capture a main design theme of a shape of the bottle as function of a highest score for each design theme.
14 . A machine-readable storage device having instructions for execution by a processor of a machine to cause the processor to perform operations to perform a method, the operations comprising:
receiving an image that includes a type of object;
segmenting the image of the object into multiple segments, including multiple body part shape segments, via a trained segmentation machine learning model; and
inputting the segments into multiple different attribute extraction models to extract different types of attributes from each of the multiple segments.
15 . The device of claim 14 wherein the trained segmentation machine learning model has been trained on multiple images of the type of object that are labeled with multiple tagged segments identified by bounding boxes.
16 . The device of claim 15 wherein the trained segmentation machine learning model comprises a mask-recurrent convolutional neural network having classes corresponding to the multiple segments of the object.
17 . The device of claim 15 wherein the type of object comprises a bottle and wherein tags comprise classes including bottle, neck, shoulder, body, top, and logo, wherein the top class includes multiple different top classes and wherein the logo class includes multiple different logo classes, and wherein the attribute extraction models include a shape attribute extraction model and a color attribute extraction model.
18 . The device of claim 14 wherein the attribute extraction models include a shape attribute extraction model and a color attribute extraction model, wherein the color attribute extraction model generates image embedding on each segment of the image and determines distance differences between the image embeddings and target color text embeddings in the same embedding space, to output target colors.
19 . The device of claim 18 wherein the color attribute extraction model is an unsupervised model and wherein a distance difference is compared to a threshold to determine that a color is present in a salient region corresponding to one of the segments and wherein the shape attribute model utilizes ratios of measurements of bonding boxes corresponding to segments to generate shape descriptions.
20 . A device comprising:
a processor; and
a memory device coupled to the processor and having a program stored thereon for execution by the processor to perform operations comprising:
receiving an image that includes a type of object;
segmenting the image of the object into multiple segments, including multiple body part shape segments, via a trained segmentation machine learning model; and
inputting the segments into multiple different attribute extraction models to extract different types of attributes from each of the multiple segments.