Synthetic Image Generation for Classifier Training

The patent creates synthetic digital images that copy the essential visual characteristics of real-world objects without requiring actual photographs. These synthetic images are generated through computer graphics and rendering techniques, replicating object appearances, lighting conditions, and environmental contexts. This allows the training process to use unlimited synthetic copies instead of time-consuming real image capture and annotation

The system performs preliminary actions by pre-defining object properties, physical characteristics, and environmental parameters in a virtual environment before generating the synthetic images. This preliminary setup includes specifying object geometries, material properties, lighting configurations, and camera parameters, which are then used to efficiently generate training images without requiring actual object photography sessions

The system makes the training data generation self-service by automatically creating synthetic images with embedded ground truth information. The synthetic image generation process inherently knows the correct labels, object boundaries, and properties since these are defined by the virtual scene parameters. This eliminates the need for human annotators to manually label images, as the system self-generates both the training images and their corresponding annotations simultaneously

The patent introduces a virtual scene description as an intermediary between the training data requirements and the final synthetic images. This intermediary layer uses structured representations of objects, environments, and camera settings to generate both visual content and annotations consistently. The virtual scene acts as a mediator that translates high-level object definitions into detailed synthetic images with automatic labeling information