Abstract: Where an event is determined to have occurred at a location within a vicinity of a plurality of actors, imaging data captured using cameras having the location is processed using one or more machine learning systems or techniques operating on the cameras to determine which of the actors is most likely associated with the event. For each relevant pixel of each image captured by a camera, the camera returns a set of vectors extending to pixels of body parts of actors who are most likely to have been involved with an event occurring at the relevant pixel, along with a measure of confidence in the respective vectors. A server receives the vectors from the cameras, determines which of the images depicted the event in a favorable view, based at least in part on the quality of such images, and selects one of the actors as associated with the event accordingly.

Abstract: Described is a system and method for collecting item and user information and utilizing that information to determine materials handling facility patterns, item trends and inventory location trends. User monitoring data for users located in a materials handling facility may include information identifying inventory locations approached by users, gaze directions of users, user dwell times, an identification of items picked by users, an identification of items placed by users, and/or in-transit times for items. This information may be aggregated and processed to determine materials handling facility patterns, item trends and/or inventory location trends.

Abstract: A model management system performs error analysis on results predicted by a machine learning model. The model management system identifies an incorrectly classified image outputted from a machine learning model and identifies using the Neural Template Matching (NTM) algorithm, an additional image correlated to the selected image. The system outputs correlated images based on a given image and a selection by a user through a user interface of a region of interest (ROI) of the given image. The region is defined by a bounding polygon input and the correlated images include features correlated to the features within the ROI. The system prompts a task associated with the additional image. The system receives a response that includes an indication that the additional image is incorrectly labeled and including a replacement label and instruct that the machine learning model be retrained using an updated training dataset that includes the replacement label.

Abstract: Described is a multiple-camera system and process for determining an item involved in an event. For example, when a user picks an item or places an item at an inventory location, image information for the item may be obtained and processed to identify the item involved in the event and associate that item with the user.

Abstract: Where an event is determined to have occurred at a location within a vicinity of a plurality of actors, imaging data captured using cameras having the location is processed using one or more machine learning systems or techniques operating on the cameras to determine which of the actors is most likely associated with the event. For each relevant pixel of each image captured by a camera, the camera returns a set of vectors extending to pixels of body parts of actors who are most likely to have been involved with an event occurring at the relevant pixel, along with a measure of confidence in the respective vectors. A server receives the vectors from the cameras, determines which of the images depicted the event in a favorable view, based at least in part on the quality of such images, and selects one of the actors as associated with the event accordingly.

Abstract: Where an event is determined to have occurred at a location within a vicinity of a plurality of actors, imaging data captured using cameras having the location is processed using one or more machine learning systems or techniques operating on the cameras to determine which of the actors is most likely associated with the event. For each relevant pixel of each image captured by a camera, the camera returns a set of vectors extending to pixels of body parts of actors who are most likely to have been involved with an event occurring at the relevant pixel, along with a measure of confidence in the respective vectors. A server receives the vectors from the cameras, determines which of the images depicted the event in a favorable view, based at least in part on the quality of such images, and selects one of the actors as associated with the event accordingly.

Abstract: Where an event is determined to have occurred at a location within a vicinity of a plurality of actors, imaging data captured using cameras having the location is processed using one or more machine learning systems or techniques operating on the cameras to determine which of the actors is most likely associated with the event. For each relevant pixel of each image captured by a camera, the camera returns a set of vectors extending to pixels of body parts of actors who are most likely to have been involved with an event occurring at the relevant pixel, along with a measure of confidence in the respective vectors. A server receives the sets of vectors from the cameras, determines which of the images depicted the event in a favorable view, based at least in part on the quality of such images, and selects one of the actors as associated with the event accordingly.

Abstract: A model management system adaptively refines a training dataset for more effective visual inspection. The system trains a machine learning model using the initial training dataset and sends the trained model to a client for deployment. The deployment process generates outputs that are sent back to the system. The system determines that performance of predictions for noisy data points are inadequate and determines a cause of failure based on a mapping of the noisy data point to a distribution generated for the training dataset across multiple dimensions. The system determines a cause of failure based on an attribute of the noisy datapoint that deviates from the distribution of the training dataset and performs refinement towards the training dataset based on the identified cause of failure. The system retrains the machine learning model with the refined training dataset and sends the retrained machine learning model back to the client for re-deployment.

Abstract: Sensors in a facility generate sensor data associated with a region of the facility, which can be used to determine a 3D location of an object in the facility. Some sensors may sense overlapping regions of the facility. For example, a first sensor may generate data associated with a first region of the facility, while a second sensor may generate data associated with a second region of the facility that partially overlaps the first region. Sensors may fail at times as determined from sensor output data or status data. In response to identifying a failed sensor, an undetected region corresponding to the failed sensor is identified, as well as a substitute sensor that partially senses the undetected region. Sensor data from the substitute sensor, such as 2D data, is acquired and used to estimate a 3D location of an object in the undetected region.

Abstract: Described are systems and techniques to generate data for display to present visualizations of data acquired from sensors in a facility. The data visualizations may be used to develop, configure, administer, or otherwise support operation of the facility. In one implementation, the visualization may include a view incorporating aggregated images acquired from multiple cameras, depth data, tracking information about objects in the facility, and so forth. An analyst may use the data visualization to determine occurrence of an action in the facility such as a pick of an item, place of an item, what item was involved with an action, what user was involved with the action, and so forth. Based on the information presented by the data visualization, changes may be made to data processing parameters.

Abstract: A user may pick an item from a first inventory location, such as in a lane on a shelf, and may return it another location that is assigned to another type of item. Described are techniques to generate tidiness data that is indicative of whether an item has been returned to an inventory location assigned to that type of item. As items are taken, information about the type of item taken and its weight are stored. When an increase in weight at a lane indicates a return of an item to the lane, the weight of the return is compared to the stored weight of the items previously taken by a user. If the weights correspond to within a threshold value, the type of item associated with the stored weight is deemed to be returned and tidiness data indicative of a tidy return of the item to its appointed lane may be generated.

Abstract: Described is a multiple-camera system and process for determining an item involved in an event. For example, when a user picks an item or places an item at an inventory location, image information for the item may be obtained and processed to identify the item involved in the event and associate that item with the user.

Abstract: Described is a multiple-camera system for use in capturing images of users within a materials handling facility and processing those images to monitor the movement of users. For large materials handling facilities, a large number of cameras may be required to monitor the facility. Processing of the data generated from a large number of cameras becomes difficult. The implementations described herein include a hierarchy that allows image data from any number of cameras within a materials handling facility to be processed without substantially increasing the processing time needed or sacrificing processing capabilities.

Abstract: During operation, a facility may utilize many sensors, such as cameras, to generate sensor data. The sensor data may be processed by an inventory management system to track objects, determine the occurrence of events at the facility, and so forth. At any given time, some of these sensors may fail to provide timely data, may fail to provide any data, may generate inaccurate data, and so forth. Described are techniques to determine failure of sensors and adjust operation of the inventory management system to maintain operability during sensor failure.

Abstract: Described is a multiple-camera system and process for detecting a user within a materials handling facility and tracking a position of the user as the user moves through the materials handling facility. In one implementation, a plurality of depth sensing cameras are positioned above a surface of the materials handling facility and oriented to obtain an overhead view of the surface of the materials handling facility, along with any objects (e.g., users) on the surface of the materials handling facility. The depth information from the cameras may be utilized to detect objects on the surface of the materials handling facility, track a movement of those objects and determine if those objects are users.

Abstract: Described is a multiple-camera system and process for disambiguating between multiple users and identifying which of the multiple users performed an event. For example, when an event is detected, user patterns near the location of the event are determined, along with touch points at the location of the event. User pattern orientation and/or arm trajectories between the event location and the user patterns may be determined and processed to disambiguate between multiple users and determine which user pattern is involved in the event.

Abstract: Described is a multiple-camera system and process for identifying a user that performed an event and associating that user with the event. For example, when an event is detected, user patterns near the location of the event are determined, along with touch points at the location of the event. User pattern orientation and/or arm trajectories between the event location and the user pattern may be determined and processed to link the user pattern to the event, thereby confirming the association between the event and the user.

Abstract: Described is a multiple-camera system and process for determining an item involved in an event. For example, when a user picks an item or places an item at an inventory location, image information for the item may be obtained and processed to identify the item involved in the event and associate that item with the user.

Abstract: Frames of sensor data may be obtained from many sensors arranged throughout a facility. These frames may be time synchronized to support further processing. For example, frames containing image data obtained at about the same time from many cameras within the facility may be used to create an aggregate or “stitched” view of the facility at that time. The synchronization may involve storing the frames from several sensors in buffers. A time window may be specified and used in conjunction with timestamps of the frames to select a set of sensor data from the buffers that are deemed to be synchronized data. The synchronized data may then be used for further processing.

Abstract: Described is a multiple-camera system for use in capturing images of users within a materials handling facility and processing those images to monitor the movement of users. For large materials handling facilities, a large number of cameras may be required to monitor the facility. Processing of the data generated from a large number of cameras becomes difficult. The implementations described herein include a hierarchy that allows image data from any number of cameras within a materials handling facility to be processed without substantially increasing the processing time needed or sacrificing processing capabilities.