We present a unique high-diversity synthetic keypoint perception 3D dataset, SKoPe3D, which includes 33 keypoint annotations for each vehicle and is collected from a roadside view using the CARLA simulator. This feature makes SKoPe3D distinct from previously released traffic monitoring datasets. Additionally, as the dataset is simulated, it can be extended using publicly released data extension scripts. Furthermore, SKoPe3D cannot be utilized for illegal surveillance as it does not include any sensitive information such as license plates, human faces, roads, and buildings. Traffic monitoring datasets with keypoint annotations and 3D/2D annotations are valuable for traffic analysis, autonomous driving, and 3D scene reconstruction, and have the potential for use in a wide variety of studies.
Moreover, we present an evaluation of a classic keypoint detector on the dataset, Keypoint R-CNN. Performance was measured using widely accepted metrics, such as percentage of correct keypoints (PCK) and precision-recall, and the experiments demonstrated the real value of such a synthetic dataset as it achieved good prediction accuracies.
The study represents a step towards bridging the gap between synthetic and real-world data, as training a keypoint detector on a synthetic dataset allows the model to generalize on real-world images to some extent. This demonstrates the knowledge-transferability between synthetic and real-world data and highlights the value of SKoPe3D.