Synthetic-to-Real Transfer with Frozen DINOv3 Features

A computer vision model is trained for card detection and segmentation using only procedurally generated synthetic data. The approach leverages a pre-trained DINOv3 model with a frozen backbone to extract rich features, while task-specific heads are trained on the synthetic dataset. This dataset includes perfect ground truth labels like bounding boxes and segmentation masks by design. The project's goal is to evaluate how well the model generalizes to detect cards in real-world photographs, demonstrating a synthetic-to-real transfer learning technique.