Data Generation

Drivence introduces a physics-aware multi-sensor simulation module, extended from MultiTest, to provide a lightweight, physically grounded simulation method that efficiently generates temporally coherent and modality-consistent multi-sensor data. Drivence’s enhanced module offers three key improvements:

(1) It enables the simultaneous insertion of multiple objects, significantly enhancing the efficiency of test case generation.

(2) Unlike MultiTest, which handles only occlusion cases where the inserted object blocks the background, Drivence’s module can also manage scenarios where the inserted object is occluded by background obstacles.

(3) More advanced visual effects are added to the generated images, such as realistic car shadows and dynamic lighting.

The detail of data generation

(1) The integration of virtual sensor outputs with the original data. Drivence simulator implements a full camera and LiDAR synthesis pipeline that, given a specified insertion configuration, automatically produces fully integrated sensor outputs. Specifically, Drivence first employs virtual sensors to render the given object instance into modality-consistent object-level images and point clouds. These sensors are configured with real-world parameters to ensure physical plausibility and incorporate lighting and shadow effects during rendering, thereby enhancing the visual realism of the generated data. Next, the object-level images and point clouds are integrated into the original scene, during which occlusion handling is performed and the portions occluded by the inserted objects are removed from the original scene. For image data, object-level images are integrated into the original images on a pixel-wise basis according to the occlusion relationships and insertion locations. For point cloud data, points occluded by the inserted objects are removed from the original scene, and the generated points are subsequently merged with the remaining original point cloud.

(2) The construction of the object model database. The object model database is constructed following the same process as in MultiTest, where ShapeNet car models are filtered based on category and extent of damage, retaining only high-quality models for database construction. Drivence has explicitly clarified these implementation details in the revised manuscript.

(3) The maintenance of globally consistent lighting and shadows. To ensure consistency of lighting and shadows across generated sequences, Drivence utilizes the Blender API to implement a globally defined sunlight-based illumination setup. Specifically, a fixed directional sunlight source provided by Blender is applied throughout the rendering process to emulate real-world outdoor lighting conditions. This guarantees that all inserted objects within a sequence share a consistent lighting direction and shadow-casting behavior, thereby achieving coherent lighting and shadows across all frames.

Visual Comparison

Multitest

Drivence

keyframe

Frst, Multitest (right side) lacks lighting and shadow effects, resulting in flat and less realistic images. More critically, it fails to handle occlusions correctly: inserted NPCs remain fully visible even when they should be partially or fully occluded by original objects in the scene, both in the image and in the point cloud. This violates basic physical realism and may lead to perception errors that lack real-world relevance.