Segmented Vehicle Image Data Generation via 3D Model Inversion

The patent inverts the conventional approach by not annotating 2D objects directly from distorted perspective images, but rather generating 2D object information from 3D model data. This inversion eliminates perspective distortion and occlusion issues, as 3D models provide accurate geometric representations that can be projected to 2D without the quality degradation that occurs when annotating directly from distorted images.

The patent introduces 3D model data as an intermediary between the vehicle and the 2D object annotation. Instead of directly extracting 2D annotations from distorted images, the system uses 3D models as a mediator that provides accurate spatial and geometric information, which is then used to generate reliable 2D object annotations for training data.

The patent performs preliminary action by pre-generating accurate 2D object information from 3D models before the actual segmentation process. This pre-processing step creates clean, distortion-free 2D annotations in advance, which can then be used during real-time operation without requiring complex distortion correction algorithms, thus simplifying the segmentation process while maintaining speed.

The patent creates a simplified copy of the 3D world in 2D space by generating accurate 2D projections from 3D models. This 2D copy contains all necessary geometric information without perspective distortion, serving as a simplified representation that can be processed efficiently during real-time segmentation while avoiding the complexity of working directly with distorted perspective images.