Seiko Epson Corporation

Dibyendu Mukherjee

Irina Kezele

Mikhail Brusnitsyn

Mirza Tahir Ahmed

Jia Li

Jie Wang

Bowen Chen

Yan He

Chen Xuan Cui

Juhan Bae

A non-transitory computer readable medium embodies instructions that cause one or more processors to perform a method. The method includes selecting a 3D model corresponding to an object, and generating a first 2D image of the 3D model in a first pose. The method further includes generating a second 2D image of the 3D model in the first pose, the second 2D image having a different texture on the 3D model than the first 2D image. The method further includes, using an algorithm, determining a first location of a first feature on the 3D model in the first 2D image and a second location of a second feature on the 3D model in the second 2D image. The method further includes calculating a difference based on the first location and the second location. The method further includes adjusting parameters representing the algorithm based on the calculated difference.

Domain adaptation losses

A non-transitory computer readable medium storing instructions to cause one or more processors to acquire an image data sequence containing images of an object in a scene along time and track a pose of the object through an object pose tracking algorithm. The processor may further, acquire a first pose of the object in a first image of the image data sequence, the first pose being a result of tracking a pose of the object through an object pose tracking algorithm, verify the first pose, extract 2D features of the object from the first image when the first pose is verified, and store a training dataset containing the extracted 2D features and the corresponding verified first pose in the one or more memories or other one or more memories.

Online learning for 3D pose estimation using simplified constellations

A non-transitory computer readable medium embodies instructions that cause one or more processors to perform a method. The method includes selecting a 3D model corresponding to an object. The method further includes generating domain-adapted images of the 3D model, the domain-adapted images representing the 3D model at corresponding poses. The method further includes acquiring 2D projections of 3D points on a 3D bounding box defined around the 3D model at the corresponding poses. The method further includes training an algorithm model to learn correspondences between the generated images and the corresponding 2D projections. The method further includes storing, in a memory, parameters representing the algorithm model.

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

A non-transitory computer readable medium embodies instructions that cause one or more processors to perform a method. The method includes: (A) receiving, in one or more memories, a 3D model corresponding to an object, and (B) setting a depth sensor characteristic data set for a depth sensor for use in detecting a pose of the object in a real scene. The method also includes (C) generating blurred 2.5D representation data of the 3D model for at least one view around the 3D model based on the 3D model and the depth sensor characteristic data set, to generate, on the basis of the 2.5D representation data, training data for training an object detection algorithm, and (D) storing the training data in one or more memories.

Methods and systems for shape based training for an object detection algorithm

A head-mounted display, a method, and a non-transitory computer readable medium are provided. An embodiment of a method for obtaining training sample views of an object includes the step of storing, in a memory, multiple views of an object. The method also includes the step of deriving similarity scores between adjacent views and then a sampling density is varied based on the similarity scores.

Adaptive sampling of training views

A method includes acquiring, from a camera, an image data sequence of a real object in a real scene and performing a first template-matching on an image frame in the image data sequence using intensity-related data sets stored in one or more memories to generate response maps. The intensity-related data sets represent an intensity distribution of a reference object from respective viewpoints. The reference object corresponds to the real object. A candidate region of interest is determined for the real object in the image frame based on the response maps, and second template-matching is performed on the candidate region of interest using shape-related feature data sets stored in one or more memories to derive a pose of the real object. The shape-related feature data sets represent edge information of the reference object from the respective viewpoints.

Growing community of inventors

Scarborough, Canada

Dibyendu Mukherjee