Synthetic Image Augmentation for Autonomous Perception Training

The patent uses synthetic image generation to create copies of rare objects and agents that cannot be easily captured in real-world driving. By rendering 3D models of traffic hazards, pedestrians, bicycles, and other rare objects into synthetic images, the system supplements the limited real-world dataset with diverse synthetic counterparts, thereby increasing overall data diversity without requiring extended real-world collection periods.

The system performs preliminary data preparation by pre-rendering synthetic images of rare objects under various conditions (lighting, weather, angles) before they are needed for training. This advance preparation ensures that diverse object representations are already available when the model training begins, eliminating the need to wait for rare real-world occurrences during data collection.

The patent merges real-world images with synthetic images of rare objects to create a hybrid training dataset. The real images provide authentic background scenes and common objects, while the synthetic images contribute diverse representations of rare traffic hazards, pedestrians, and cyclists. This combination allows the model to learn from both actual driving scenarios and supplemented rare object examples, improving overall training reliability.

The system uses an intermediary synthetic data generation pipeline that bridges the gap between real-world data limitations and model training requirements. The synthetic image generation process acts as a mediator, translating 3D models of rare objects into realistic 2D images that can be seamlessly integrated with real-world photographs, thereby enabling the model to learn about rare objects without direct real-world exposure.