Methods and systems for training an object detection algorithm using synthetic images
AssigneeSeiko Epson Corporation
InventorDibyendu MUKHERJEE
Life of this patent
Estimating the term…
Abstract
A non-transitory computer readable medium embodies instructions that cause one or more processors to perform a method for training an object detection algorithm. The method includes: (a) selecting a 3D model corresponding to an object; (b) acquiring images of the 3D model, the images being obtained by rendering the 3D model at respective poses; (c) acquiring 2D projections of 3D points on the 3D model at the respective poses; and (d) storing, in a memory, an association between the acquired 2D projections and the respective poses.
Claims21
A non-transitory computer readable medium that embodies instructions that cause one or more processors to perform a method for training an object detection algorithm, the method comprising:
- (a) selecting a 3D model corresponding to an object;
- (b) acquiring images of the 3D model, the images being obtained by rendering the 3D model at respective poses;
- (c) acquiring 2D coordinate values representing 2D projections of 3D points on a 3D surface forming at least a part of the 3D model at the respective poses;
- (e1) training an algorithm model to learn correspondences between the acquired images and the respective 2D coordinate values after step (c);
- and
- (d) storing, in a memory, parameters representing the algorithm model and an association between the acquired 2D coordinate values and the respective poses.
(canceled)
A non-transitory computer readable medium that embodies instructions that cause one or more processors to perform a method for an object detection algorithm, the method comprising:
- (a) acquiring, from a camera, an image containing an object in a scene;
- (b) deriving 2D coordinate values by using a trained algorithm model with the image as input, the 2D coordinate values representing 2D projections of 3D points on a 3D surface forming at least a part of a 3D model corresponding to the object;
- and
- (c) deriving a pose of the object based on the derived 2D points,
- wherein the algorithm model is trained to learn correspondences between (i) a synthetic image of the 3D model at a corresponding pose and (ii) corresponding 2D coordinate values representing 2D projections of 3D points on the 3D surface forming at least the part of the 3D model at the corresponding pose.
Dependencies are inferred from the claim text.