SDOA: A Synthetic Dataset for Occlusion-Aware and Amodal Perception in Autonomous Driving

Humans can easily infer the complete appearance of partially occluded objects. However, current computer vision algorithms perform poorly in occluded scenes. Due to the high cost, low efficiency, and potential safety risks of real-world data collection, existing datasets lack sufficient annotations for occluded scenes, which severely restricts the application of amodal perception in autonomous driving. To address this issue, this study proposes an efficient simulation-based occluded scene generation method for autonomous driving. It constructs a Synthetic Dataset for Occlusion-Aware and Amodal Perception, namely SDOA (Synthetic Dataset for Occlusion-Aware and Amodal Perception), to meet the requirements of amodal perception tasks. This method adopts a concise parametric scene description and follows a “random first, then adjust” strategy. First, it generates an initial scene that satisfies basic constraints (e.g., objects are on the road, non-overlapping, and visible) using a dynamic generation strategy. Then, it optimizes the scene through an occlusion adjustment strategy to generate scene instances with a specific occlusion ratio. The simulator's sensor data is used to generate ground-truth labels such as modal masks, amodal masks, and occlusion ratios, providing essential data for training and evaluating perception models. Our SDOA dataset and its generation method can offer crucial data for autonomous driving amodal perception, significantly promote the research and application of perception algorithms in occluded scenes, and lay a solid foundation for optimizing the safety of autonomous driving systems.