The vehicle–infrastructure–air (VIA) collaborative perception dataset, collected through the simulation platform, contains multi-view images and point cloud data from connected autonomous vehicles, roadside infrastructure, and aerial platforms. The dataset provides annotated ground-truth bounding boxes, sensor calibration parameters, and temporal sequence information.
Participants are required to write the inference results for each frame in the test set into a CSV file (see example below) following the specified format. The submitted CSV will be compared with the ground-truth (GT) data in the system to compute the final score.
Submission Format Instructions
The CSV file must contain only two columns:

First column: ID—the sequential index of each test frame (a total of 2170 frames).


Second column: all vehicle detection results for that frame (one or more targets).

For each frame, the predicted bounding boxes from the object detection model should be listed in order within the second column.
Each predicted bounding box must include the following values in the specified order:
vehicle score center_x center_y center_z length, width, height, rotation_x, rotation_y, rotation_z

where each item represents:

vehicle—object category
score—confidence score
center_x, center_y, center_z — 3D bounding box center coordinates
length, width, height—bounding box dimensions
rotation_x, rotation_y, rotation_z—roll, yaw, and pitch angles

⚠️ Important:
The CSV file must contain exactly two columns.
If there are multiple targets in a frame, all should be listed sequentially within the same cell of the second column.
Any formatting errors in the CSV file will result in failure to obtain a score.