Seiko Epson Corporation

Ivo Moravec

Syed Alimul Huda

Arash Abadpour

Jie Wang

Michael Joseph Mannion

Guoyi Fu

Yang Yang

Bogdan Matei

Eunice Poon

Irina Kezele

Ian Clarke

Mikhail Brusnitsyn

Sujay Sukumaran

Yury Yakubovich

Lihui Chen

Vivek Mogalapalli

Dan Rico

A method includes: (A) receiving a selection of a 3D model stored in one or more memories, the 3D model corresponding to an object and (B) setting a camera parameter set for a camera for use in detecting a pose of the object in a real scene. The method also includes (C) generating at least one 2D synthetic image based at least on the camera parameter set by rendering the 3D model in a view range for generating training data.

Methods and systems for training an object detection algorithm using synthetic images

A method may include the following steps: acquiring, from a camera, an image frame; acquiring, from an inertial sensor, a sensor data sequence; tracking a first pose of an object in a real scene based at least on the image frame; deriving a sensor pose of an inertial sensor based on the sensor data sequence; determining whether the first pose is lost; retrieving from one or more memories, or generating from a 3D model stored in one or more memories, a training template corresponding to a view that is based on the sensor pose obtained on or after the first pose is lost; and deriving a second pose of the object using the training template.

Object detection and tracking

A non-transitory computer readable medium embodies instructions that cause one or more processors to perform a method. The method includes: (A) receiving a selection of a 3D model stored in one or more memories, the 3D model corresponding to an object, and (B) setting a camera parameter set for a camera for use in detecting a pose of the object in a real scene. The method also includes (C) receiving a selection of data representing a view range, (D) generating at least one 2D synthetic image based on the camera parameter set by rendering the 3D model in the view range, (E) generating training data using the at least one 2D synthetic image to train an object detection algorithm, and (F) storing the generated training data in one or more memories.

Multiple Holocam Orbs observe a real-life environment and generate an artificial reality representation of the real-life environment. Depth image data is cleansed of error due to LED shadow by identifying the edge of a foreground object in an (near infrared light) intensity image, identifying an edge in a depth image, and taking the difference between the start of both edges. Depth data error due to parallax is identified noting when associated text data in a given pixel row that is progressing in a given row direction (left-to-right or right-to-left) reverses order. Sound sources are identified by comparing results of a blind audio source localization algorithm, with the spatial 3D model provided by the Holocam Orb. Sound sources that corresponding to identifying 3D objects are associated together. Additionally, types of data supported by a standard movie data container, such as an MPEG container, is expanding to incorporate free viewpoint data (FVD) model data. This is done by inserting FVD data of different individual 3D objects at different sample rates into a single video stream. Each 3D object is separately identified by a separately assigned ID.

Encoding free view point data in movie data container

Cleaning of depth data by elimination of artifacts caused by shadows and parallax

A three-dimensional image processing apparatus includes: an obtainment unit that obtains range image data from each of a plurality of range image generation devices and obtains visible light image data from each of a plurality of visible light image generation devices; a model generation unit that generates three-dimensional model data expressing a target contained in a scene based on a plurality of pieces of the range image data; a setting unit that sets a point of view for the scene; and a rendering unit that selects one of the pieces of the visible light image data in accordance with the set point of view and renders a region corresponding to the surface of the target based on the selected visible light image data.

Three-dimensional image processing apparatus and three-dimensional image processing system

Aspects of the present invention comprise holocam systems and methods that enable the capture and streaming of scenes. In embodiments, multiple image capture devices, which may be referred to as 'orbs,' are used to capture images of a scene from different vantage points or frames of reference. In embodiments, each orb captures three-dimensional (3D) information, which is preferably in the form of a depth map and visible images (such as stereo image pairs and regular images). Aspects of the present invention also include mechanisms by which data captured by two or more orbs may be combined to create one composite 3D model of the scene. A viewer may then, in embodiments, use the 3D model to generate a view from a different frame of reference than was originally created by any single orb.

Holocam systems and methods

An adequate solution for computer vision applications is arrived at more efficiently and, with more automation, enables users with limited or no special image processing and pattern recognition knowledge to create reliable vision systems for their applications. Computer rendering of CAD models is used to automate the dataset acquisition process and labeling process. In order to speed up the training data preparation while maintaining the data quality, a number of processed samples are generated from one or a few seed images.

Method and apparatus for improved training of object detecting system

A pose of an object is estimated from an from an input image and an object pose estimation is then stored by: inputting an image containing an object; creating a binary mask of the input image; extracting a set of singlets from the binary mask of the input image, each singlet representing points in an inner and outer contour of the object in the input image; connecting the set of singlets into a mesh represented as a duplex matrix; comparing two duplex matrices to produce a set of candidate poses; and producing an object pose estimate, and storing the object pose estimate. The estimated pose of the object is refined by: inputting an image of an object in an estimated pose, a model of the object, and parameters of a camera used to take the image of the object in the estimated pose; projecting the model of the object into a virtual image of the object using the parameters of the camera and initial pose parameters to obtain a binary mask image and image depth information; and updating the initial pose parameters to new pose parameters using the binary mask image and image depth information and updating the new pose parameters iteratively to minimize an energy function or until a maximum number of iterations is reached.

Growing community of inventors

Richmond Hill, Canada

Ivo Moravec