Generative Transformer Model Retraining for Vehicle Object Detection

The patent applies preliminary action by pre-training the generative deep learning model on large-scale natural language data before fine-tuning it for specific object detection tasks. This pre-training phase prepares the model with general language understanding and patterns, so that when it is later fine-tuned with smaller, task-specific datasets, it achieves high performance with reduced training effort and annotation requirements.

The patent applies self-service by enabling the generative model to automatically generate training data and annotations through its language understanding capabilities. Instead of requiring extensive manual or automated annotation processes, the model can generate synthetic training examples with automatic annotations, thereby improving training data quality while eliminating the resource-intensive annotation process.

The patent applies parameter changes by fine-tuning the pre-trained generative model's parameters through retraining on smaller, task-specific datasets containing vehicle environment images and annotations. This process adjusts the model's parameters to be optimized for specific detection tasks such as detecting pedestrians, vehicles, or obstacles in automotive contexts, thereby reducing training effort while achieving task-specific adaptability.