3D-aware image search
Systems and methods for performing image search are described. An image search method may include generating a feature vector for each of a plurality of stored images using a machine learning model trained using a rotation loss term, receiving a search query comprising a search image with object having an orientation, generating a query feature vector for the search image using the machine learning model, wherein the query feature vector is based at least in part on the orientation, comparing the query feature vector to the feature vector for each of the plurality of stored images, and selecting at least one stored image of the plurality of stored images based on the comparison, wherein the at least one stored image comprises a similar orientation to the orientation of the object in the search image.
1. A method for image searching, comprising:
generating a feature vector for each of a plurality of stored images using a machine learning model trained to generate representations of a three-dimensional (3D) object orientations;
receiving a search query comprising a search image with an object having a 3D orientation;
generating a query feature vector for the search image using the machine learning model, wherein the query feature vector is based at least in part on the 3D orientation;
comparing the query feature vector to the feature vector for each of the plurality of stored images; and
selecting at least one stored image of the plurality of stored images based on the comparison, wherein the at least one stored image comprises a similar orientation to the 3D orientation of the object in the search image.
2. The method of claim 1 , wherein:
the machine learning model is trained using a rotation loss term.
3. The method of claim 2 , wherein:
the machine learning model trained using an object synthesis loss term or an image ranking loss term.
4. The method of claim 1 , further comprising:
providing the at least one stored image in response to the search query.
5. The method of claim 1 , further comprising:
determining that a search parameter of the search query indicates a search for objects based on the 3D orientation, wherein the query feature vector is generated based on the search parameter.
6. The method of claim 1 , wherein:
the query feature vector further represents semantic information corresponding to the object.
7. The method of claim 1 , wherein:
comparing the query feature vector comprises calculating a Euclidean distance or a cosine distance between the feature vector and the feature vector for each of the plurality of stored images.
8. A system for image processing, comprising:
an encoder configured to generate a feature vector based on an image of an object, wherein the feature vector represents object synthesis information and rotation information, and wherein the encoder is trained using a loss function comprising a rotation loss term that measures whether an output of the encoder represents the rotation information;
a comparison component configured to determine a similarity between two feature vectors;
a query component configured to receive a search query comprising the image of the object; and
a results component configured to select one or more additional images based on the similarity, and to provide the one or more additional images in response to the search query.
9. The system of claim 8 , wherein:
the encoder comprises a convolutional neural network (CNN) layer and a multilayer perceptron (MLP) layer.
10. The system of claim 8 , wherein:
the loss function comprises a synthesis loss term that measures whether the object can be synthesized based on the feature vector.
11. The system of claim 8 , wherein:
the rotation loss term is based on a measure of a difference in a pose of the object compared to a pose of at least one additional object.
12. The system of claim 8 , wherein:
the loss function further comprises a ranking loss term or a classification loss term.
13. The system of claim 12 , wherein:
the ranking loss term is based on selecting two images from a set of three images based on relevance.
14. The system of claim 12 , wherein:
the classification loss term is based on object classification accuracy.
15. A method of training an artificial neural network (ANN) for image processing, comprising:
identifying a training set, wherein each element in the training set includes an image of an object and 3D orientation information for the object;
generating a feature vector for each element in the training set;
calculating a loss value using a loss function comprising a rotation loss term that measures whether an output of the ANN represents rotation information; and
updating the ANN based at least in part on the loss value.
16. The method of claim 15 , wherein:
the loss value is calculated, and the ANN is updated, during each of a plurality of training iterations.
17. The method of claim 15 , further comprising:
generating a synthesized version of the object based on the feature vector; and
calculating an object synthesis loss term based on a comparison between the object and the synthesized version of the object.
18. The method of claim 15 , wherein:
the loss function further comprises a ranking loss term or a classification loss term.
19. The method of claim 18 , further comprising:
identifying two additional images; and
selecting a most relevant image from the two additional images, wherein the ranking loss term is based on the selection.
20. The method of claim 19 , wherein:
the ranking loss is based on one or more parameters from a set consisting of comparing pose similarity, object similarity, and category similarity.