A Novel Synthetic Dataset for Broadcast Motorsports Scene Understanding

The paper introduces a foundational approach to motorsports scene understanding by investigating the role of synthetic data generation in advancing scene understanding for high-speed broadcast scenarios. Utilizing the CARLA simulation environment, the study constructs a high-fidelity dataset incorporating diverse lighting conditions, occlusions, and dynamic camera perspectives to enhance model generalization. A multi-stage data refinement pipeline is introduced to mitigate the impact of extreme occlusions and irrelevant samples while preserving the complexity of real-world challenges. Possible applications include 3D real-world understanding from a single monocular 2D image, which could open up interesting possibilities for augmented reality in broadcast media by allowing seamless integration of virtual elements, interactive graphics and dynamic visual effects, enhancing storytelling, audience engagement, and production flexibility. The efficacy of the dataset is further evaluated via transfer learning to the real-world domain, with the model pretrained on synthetic data demonstrating a significantly superior performance compared to its counterpart.