Abstract: An interactive interpretation session with respect to a first version of a machine learning model is initiated. In the session, indications of factors contributing to a prediction decision are provided, as well indications of candidate model enhancement actions. In response to received input, an enhancement action is implemented to obtain a second version of the model. The second version of the model is stored.

Abstract: Computer systems and associated methods are disclosed to implement a model development environment (MDE) that allows a team of users to perform iterative model experiments to develop machine learning (ML) media models. In embodiments, the MDE implements a media data management interface that allows users to annotate and manage training data for models. In embodiments, the MDE implements a model experimentation interface that allows users to configure and run model experiments, which include a training run and a test run of a model. In embodiments, the MDE implements a model diagnosis interface that displays the model's performance metrics and allows users to visually inspect media samples that were used during the model experiment to determine corrective actions to improve model performance for later iterations of experiments. In embodiments, the MDE allows different types of users to collaborate on a series of model experiments to build an optimal media model.

Abstract: An interactive interpretation session with respect to a first version of a machine learning model is initiated. In the session, indications of factors contributing to a prediction decision are provided, as well indications of candidate model enhancement actions. In response to received input, an enhancement action is implemented to obtain a second version of the model. The second version of the model is stored.

Abstract: Computer systems and associated methods are disclosed to implement a model development environment (MDE) that allows a team of users to perform iterative model experiments to develop machine learning (ML) media models. In embodiments, the MDE implements a media data management interface that allows users to annotate and manage training data for models. In embodiments, the MDE implements a model experimentation interface that allows users to configure and run model experiments, which include a training run and a test run of a model. In embodiments, the MDE implements a model diagnosis interface that displays the model's performance metrics and allows users to visually inspect media samples that were used during the model experiment to determine corrective actions to improve model performance for later iterations of experiments. In embodiments, the MDE allows different types of users to collaborate on a series of model experiments to build an optimal media model.

Abstract: Computer systems and associated methods are disclosed to implement a model development environment (MDE) that allows a team of users to perform iterative model experiments to develop machine learning (ML) media models. In embodiments, the MDE implements a media data management interface that allows users to annotate and manage training data for models. In embodiments, the MDE implements a model experimentation interface that allows users to configure and run model experiments, which include a training run and a test run of a model. In embodiments, the MDE implements a model diagnosis interface that displays the model's performance metrics and allows users to visually inspect media samples that were used during the model experiment to determine corrective actions to improve model performance for later iterations of experiments. In embodiments, the MDE allows different types of users to collaborate on a series of model experiments to build an optimal media model.

Abstract: Computer systems and associated methods are disclosed to implement a collaborative dataset management system (CDMS) for machine learning (ML) data. In embodiments, CDMS allows many users to create, review, and collaboratively evolve ML datasets. In embodiments, dataset owners may make their datasets available to other users on CDMS for a fee and under specified licensing conditions. CDMS users can search for other users' datasets on the system to use in their own ML tasks. CDMS users may also create child datasets from existing datasets on the system. Parent and child datasets may be linked so that changes to one dataset are provided to the other via merge requests. A dataset owner may use CDMS to review an incoming merge request using one or more audit jobs before approving the request. In this manner, CDMS provides a shared repository and collaboration system for managing high-quality datasets to power machine learning processes.

Abstract: Computer systems and associated methods are disclosed to implement a model development environment (MDE) that allows a team of users to perform iterative model experiments to develop machine learning (ML) media models. In embodiments, the MDE implements a media data management interface that allows users to annotate and manage training data for models. In embodiments, the MDE implements a model experimentation interface that allows users to configure and run model experiments, which include a training run and a test run of a model. In embodiments, the MDE implements a model diagnosis interface that displays the model's performance metrics and allows users to visually inspect media samples that were used during the model experiment to determine corrective actions to improve model performance for later iterations of experiments. In embodiments, the MDE allows different types of users to collaborate on a series of model experiments to build an optimal media model.

Abstract: A system and method that gathers and labels images for use by a machine learning algorithm for image classification is disclosed. The method includes acquiring a preview image including a portion of a scene, generating a user interface including the preview image, performing object detection on the preview image to detect a set of items, adding, to the first preview image, a set of regions of interest based on the object detection, each region of interest highlighting a location of a corresponding item in the set of items, receiving an image corresponding to the preview image, determining an object identifier associated with an item, and labeling the item in the image using the object identifier and a region of interest corresponding to the item.

Abstract: The disclosure includes a system and method for distinguishing between stock keeping units of similar appearance that vary in size. An image recognition application receives an image depicting a plurality of items, the image including a reference marker with a known physical dimension. The image recognition application performs image recognition to identify an item in the image and a region of interest for the identified image. The image recognition application further determines a pixel-to-physical dimension ratio using the dimension of a region of interest of the reference marker and the known physical dimension of the reference marker. Finally, the image recognition application determines a stock keeping unit identifier of the identified item in the image based on the pixel-to-physical dimension ratio and a dimension of the region of interest of the identified item.

Abstract: The disclosure includes a system and method for indexing synthetically modified images of a high quality image. An image recognition application receives images of a product, crops background regions from the images, scales the image based on a minimum value among width and height of the image and generates multiple image sizes, blurs the images, brightens the image and indexes the images as being associated with the product. The images can be of box-shaped packages that include four or six images or cylindrical packages that include, for example, eight images of the packages. The images can be indexed in a k-dimensional tree for faster retrieval.

Abstract: A system and method that gathers and labels images for use by a machine learning algorithm for image classification is disclosed. The method includes acquiring a preview image including a portion of a scene, generating a user interface including the preview image, performing object detection on the preview image to detect a set of items, adding, to the first preview image, a set of regions of interest based on the object detection, each region of interest highlighting a location of a corresponding item in the set of items, receiving an image corresponding to the preview image, determining an object identifier associated with an item, and labeling the item in the image using the object identifier and a region of interest corresponding to the item.

Abstract: The disclosure includes a system and method for distinguishing between stock keeping units of similar appearance that vary in size. An image recognition application receives an image including a shelving unit stocking a plurality of items, identifies each item in the image, generates a region of interest for each identified item in the image, identifies a physical dimension of a portion of region depicted in the image, determines a dimension of the region of interest for each identified item and the portion of region in pixels, determines a pixel-to-physical dimension ratio using the dimension in pixels of the portion of region and the physical dimension of the portion of region depicted in the image, and determines a stock keeping unit identifier of each identified item in the image based on the pixel-to-physical dimension ratio and the dimension of the region of interest for each identified item.

Abstract: A system and method that generates a candidate list of indexed images that potentially match an object in a query image is disclosed. The method includes receiving a query image including an object, receiving a plurality of indexed images that match the object, computing a region of interest for the object, computing an overlap between a first region of interest corresponding to a first indexed image and a second region of interest corresponding to a second indexed image, determining that the overlap between the first region of interest and the second region of interest satisfies a threshold and including the first indexed image and the second indexed image in a candidate list of indexed images.

Abstract: The disclosure includes a system and method for distinguishing objects of similar appearance that vary in color. An image recognition application receives a query image of a plurality of objects, determines a candidate list of indexed images that match an object in the query image based on luminance features, crops an image of the object from the query image to generate a cropped image, computes color features for the object, ranks the candidate list of indexed images based on the color features, and provides the candidate list of indexed images to a user.

Abstract: A system and method that computes a quality score for an index image is disclosed. The method includes receiving an index image, computing a blurriness score of the index image based on variance associated with the index image, computing an image resolution score of the index image based on an area of the index image and a threshold area, computing a feature spread score for the index image using a first plurality of features associated with the index image, computing a feature uniqueness score for the index image using a description associated with each of a second plurality of features and determining a quality score for the index image using the blurriness score, the image resolution score, the feature spread score, and the feature uniqueness score.

Abstract: The disclosure includes a system and method for distinguishing objects of similar appearance that vary in color. An image recognition application receives a query image of a plurality of objects, determines a candidate list of indexed images that match an object in the query image based on luminance features, crops an image of the object from the query image to generate a cropped image, computes color features for the object, ranks the candidate list of indexed images based on the color features, and provides the candidate list of indexed images to a user.

Abstract: A system and method that generates a candidate list of indexed images that potentially match an object in a query image is disclosed. The method includes receiving a query image including an object, receiving a plurality of indexed images that match the object, computing a region of interest for the object, computing an overlap between a first region of interest corresponding to a first indexed image and a second region of interest corresponding to a second indexed image, determining that the overlap between the first region of interest and the second region of interest satisfies a threshold and including the first indexed image and the second indexed image in a candidate list of indexed images.

Abstract: A system and method that computes a quality score for an index image is disclosed. The method includes receiving an index image, computing a blurriness score of the index image based on variance associated with the index image, computing an image resolution score of the index image based on an area of the index image and a threshold area, computing a feature spread score for the index image using a first plurality of features associated with the index image, computing a feature uniqueness score for the index image using a description associated with each of a second plurality of features and determining a quality score for the index image using the blurriness score, the image resolution score, the feature spread score, and the feature uniqueness score.

Abstract: The disclosure includes a system and method for distinguishing between items of similar appearance that vary in size. An image recognition application receives an image including a plurality of items, identifies each item in the image, generates a region of interest for each identified item in the image, determines a dimension of the region of interest for each identified item in pixels, determines a plurality of pixel-to-physical dimension ratios for each identified item, identifies a selected pixel-to-physical dimension ratio among the plurality of pixel-to-physical dimension ratios, and determines a stock keeping unit identifier of each identified item in the image based on the selected pixel-to-physical dimension ratio and the dimension of the region of interest for each identified item.

Abstract: A method for sequencing products on a retail shelf, or other objects, recognized in an image of the retail shelf includes detecting products in the received image and calculating a final sequence position for each detected product. The final sequence position is calculated by generating a region of interest for each detected product in the received image, generating a plurality of unique raster numbers for each region of interest using a raster number equation by varying parameters of the raster number equation based on pixel coordinates of each region of interest, determining a plurality of sequence positions for each region of interest using the plurality of unique raster numbers, and determining the final sequence position for each region of interest using the plurality of sequence positions for each region of interest. The disclosure also includes systems for implementing the method.