แนวคิดหลัก
Proposing Pseudo Ground Truth Augmentation (PGT-Aug) to address class imbalance in LiDAR-based object detection.
บทคัดย่อ
The content introduces PGT-Aug, a method to augment rare objects using pseudo-LiDAR point clouds generated from videos. It addresses the class imbalance problem in LiDAR datasets by leveraging miniatures and public videos. The framework involves volumetric 3D instance reconstruction, domain alignment, and point cloud augmentation. Extensive experiments on nuScenes, KITTI, and Lyft datasets demonstrate the effectiveness of PGT-Aug in improving detection performance for minority classes.
สถิติ
"Our bank has 36,960 trucks, 52,800 construction vehicles, 12,960 buses, 19,280 trailers, 25,279 motorcycles, and 4,300 bicycles."
"We trained the network for 200 epochs with a batch size of 300 sample pairs from RGB and intensity domains."
คำพูด
"Our method can effectively operate across various scenes and outperform existing approaches."
"To deal with the class imbalance problem is collecting more LiDAR data but obtaining sufficient long-tail samples is practically challenging."