Transformation Operators

This page contains details on the design of transformation operators in V2XGen's workflow. Here we provide more details about sensor parameter configuration and visualisation of entity insertion and deletion. 

Sensor Parameter Configuration

Given that the V2V4Real [1] dataset is collected using multiple vehicles outfitted with Velodyne VLP-32 LiDAR sensors, we use the same Velodyne VLP-32 LiDAR configuration used to collect the dataset to build our virtual simulator, with some parameters as follows:

Occlusion Handling

The details of the occlusion handling operation are as follows:

1. First, we obtain the minimum convex hull of the entity by projecting it according to its position in the ego and cooperative LiDAR coordinate system.

2. Next, based on the position of LIDAR and the convex hull of the entity, we could construct a vertebral-like region containing a visible region and an invisible region.

3. Finally, we process interaction with background objects in the scene at each viewpoint. For each viewpoint, we delete the point clouds occluded by the inserted entity in the scene and remove the occluded parts in the inserted entity according to the geometric relationship.

Entity Deletion

Occlusion Completion

After deleting entities’ point clouds, V2XGen leverages a physic-based occlusion-completion strategy based on virtual sensor simulation to reconstruct the previously occluded background ground and objects.

The details of the occlusion completion operation are as follows:

1. First, We leverage the multi-view virtual LiDAR sensor set to infer the invisible occlusion region of the selected entity instance and sequentially complement the ground and objects in the invisible area.

2. Next, to complete the ground, we meshify the road point cloud to reconstruct the road surface and leverage the ray casting algorithm to complete the obscured ground with realistic scan lines.

3. Then, to complete the background objects, we compute the intersection of the occluded region with the bounding boxes of all background objects to identify those that were occluded by the deleted entity.

4. Finally, we reinsert entities of the same size as the originals at the locations of each occluded object and leverage the multi-view virtual LiDAR sensor set to reconstruct the occluded object under each view.

[1] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng et al. "V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception." In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 13712-13722. 2023.