2D to 3D Label Propagation For Object Detection In Point Cloud

摘要

Object detection and classification from LiDAR point cloud is increasingly important in robotic system. For training a classifier, huge datasets with object labeling is needed. However, manually labeled data from point cloud is time-consuming and costly. We present a framework, which propagates image annotation to point cloud for making a training data. Each object point cloud is projected on the corresponding image and searches for the overlapping area within the 2D bounding box. If the occurred area is matched, the label propagates to the point cloud object. While previous works trained their classifier with 3D ground truth and tested by combining point cloud and RGB image. Our system trained and tested using only point cloud to identify objects. The comparison between using manual labeling and label propagation training data demonstrate that the label propagation can be used to train the classifier without manually ground truth with 80.22% mean average precision.