Evaluation Dataset for LiDAR-Based Object Detection Algorithms Across Varying Angular Resolutions

This dataset provides object detection results using five different LiDAR-based object detection algorithms: **PointRCNN**, **SECOND**, **Part-A²**, **PointPillars**, and **PVRCNN**. The experiments aim to determine the optimal angular resolution for LiDAR-based object detection. The point cloud data was generated in the CARLA simulator, modeled in a suburban scenario featuring 30 vehicles, 13 bicycles, and 40 pedestrians. The angular resolution in the dataset ranges from 0.1° x 0.1° (H x V) to 1.0° x 1.0°, with increments of 0.1° in each direction.

For each angular resolution, over 2000 frames of point clouds were collected, with 1600 of these frames labeled across three object classes—vehicles, pedestrians, and cyclists, for algorithm training purposes The dataset includes detection results after evaluating 1000 frames, with results recorded for the respective angular resolutions.

Each file in the dataset contains five sheets, corresponding to the five different algorithms evaluated. The data structure includes the following columns:

1. **Frame Index**: Indicates the frame number, ranging from 1 to 1000.

2. **Object Classification**: Labels objects as 1 (Vehicle), 2 (Pedestrian), or 3 (Cyclist).

3. **Confidence Score**: Represents the confidence level of the detected object in its bounding box.

4. **Number of LiDAR Points**: Indicates the count of LiDAR points within the bounding box.

5. **Bounding Box Distance**: Specifies the distance of the bounding box from the LiDAR sensor.

**This dataset has been created in the context of the Leibniz Young Investigator Grants- programmed by the Leibniz University Hannover and is funded by the Ministry of Science and Culture of Lower Saxony (MWK) Grant Nr. 11-76251-114/2022**