Dataset

"Domain-Invariant 3D Structural Convolution Network for Autonomous Driving Point Cloud Dataset"

Link:

https://share.mlmlab.synology.me/sharing/nWUDWeFaR

This dataset includes car, pedestrian, and truck three classes. Each data was produced using ground truth data of KITTI, NuScenes, and Pankyo datasets measured using actual LiDAR sensors. The structure of the dataset will be as follows:

3_classes_GT

├── kitti_3

│   ├── car

│   │   ├── test

│   │   ├── train

│   ├── pedestrian

│   │   ├── test

│   │   ├── train

│   ├── truck

│   │   ├── test

│   │   ├── train

├── nusenes_3

│   ├── car

│   │   ├── test

│   │   ├── train

│   ├── pedestrian

│   │   ├── test

│   │   ├── train

│   ├── truck

│   │   ├── test

│   │   ├── train

├── pankyo_3

│   ├── car

│   │   ├── test

│   │   ├── train

│   ├── pedestrian

│   │   ├── test

│   │   ├── train

│   ├── truck

│   │   ├── test

│   │   ├── train

Common

Ground truth data of above 120 points for car and truck classes and above 80 points for pedestrian classes were used, and ground truth data of more than 3000 points were generated by randomly sampling 3000 points among even-point indices. Thus, all data can contain up to 3000 points. All classes consist of a .ply format of 8:2 ratio of train data and test data. All datasets contain only xyz's information.

KITTI

Data from KITTI 3D detection data were used, and ground truth data generated through mmdetection3d were used.

NuScenes

Data from nuScenes 3D detection were used, and ground truth data generated through openPCD were used. NuScenes ground truth data was generated by removing the floor points in the case of data with more than 1000 points because floor points were often included.

Pankyo

Data from the Gyeonggi Data Dream portal were utilized, collated in Pangyo Zero City with 128-channel Velodyne LiDARs. This data was used to explore the applicability of LiDAR data, aiming to diversify LiDAR channels.

The modification of the data was achieved not through complex manipulation techniques but rather through a straightforward approach of using annotations. Specifically, we utilized the bounding boxes provided within the dataset to generate ground truth data for objects.

Moreover, acknowledging the inherent uncertainties within the annotations, particularly regarding the accuracy of object delineations, we implemented a simple yet effective strategy to enhance the dataset's quality. We removed the bottom 10% of the data along the Z-axis to eliminate ground points.