Attitude Estimation Accuracy Degradation Detection

The patent creates virtual images through computer graphics (CG) to copy real-world scenarios, allowing the generation of diverse training data without physically capturing each scenario. This enables comprehensive coverage of attitudes and lighting environments at lower cost while maintaining estimation accuracy.

The patent pre-generates a comprehensive dataset covering all possible attitudes and lighting environments before actual operation. This preliminary action ensures the model is trained on diverse data, improving accuracy while avoiding the need to generate data during operational phases.

The patent creates a universal dataset that covers multiple attitudes, lighting environments, and object types. This multi-functional dataset enables the model to handle various unanticipated situations reliably while keeping generation costs manageable through efficient virtual image synthesis.

The patent systematically varies parameters such as lighting conditions, object attitudes, and camera angles in virtual image generation. By changing these parameters comprehensively, the dataset becomes robust to unanticipated situations while controlling generation costs through targeted parameter exploration.

The patent uses CG-generated images as an intermediary to bridge the gap between controlled virtual environments and real-world applications. By carefully designing the CG rendering process to match real-world characteristics, the intermediary data maintains accuracy while reducing generation costs.

The patent applies different quality levels to different parts of the dataset generation process. Critical regions requiring high fidelity are rendered with greater detail, while less critical areas use optimized rendering, balancing accuracy and cost in the overall dataset.