Abstract: The disclosure notably relates to a computer-implemented method for forming a dataset configured for learning a function. The function is configured to classify 2D images according to predetermined types of 3D arrangement with respect to objects visible in the 2D images. The method comprising for each respective type of 3D arrangement, constructing 3D scenes each comprising 3D modeled objects arranged according to the respective type of 3D arrangement, generating 2D images each representing a respective 2D perspective of a respective constructed 3D scene where visible 3D modeled objects are among the 3D modeled objects of the respective constructed 3D scene which are arranged according to the respective type of 3D arrangement, and adding to the dataset training patterns each including a respective generated 2D image and information indicative of the respective type of 3D arrangement. Such a method improves 2D image classification.

Abstract: The disclosure notably relates to a computer-implemented method for forming a dataset configured for learning a function. The function is configured to classify 2D images according to predetermined types of 3D arrangement with respect to objects visible in the 2D images. The method comprising for each respective type of 3D arrangement, constructing 3D scenes each comprising 3D modeled objects arranged according to the respective type of 3D arrangement, generating 2D images each representing a respective 2D perspective of a respective constructed 3D scene where visible 3D modeled objects are among the 3D modeled objects of the respective constructed 3D scene which are arranged according to the respective type of 3D arrangement, and adding to the dataset training patterns each including a respective generated 2D image and information indicative of the respective type of 3D arrangement. Such a method improves 2D image classification.