Also, we compare the regression targets of 3d object detection networks with and without anchor boxes. Then, we present the focal loss and hard negative mining as the common methods to address the background class-imbalance in classification loss of 3d object detection networks, and smooth L1 as a robust regression loss against outliers. We start the 3d object detection background section with formally defining 3d object detection task as well as reviewing the 6 degrees of freedom to encode each predicted oriented 3d box. Next, the data format of returned lidar points is described. Then, we officially define the lidar coordinate frame which is used as the coordinate frame to represent the coordinates of returned lidar points as well as the predicted oriented 3d boxes at the output of detection networks. In the lidar point clouds section, first, we review the KITTI dataset that has become the standard benchmark for self-driving perception tasks. This blog post is divided into three main sections: lidar point clouds, 3d object detection background and 3d object detection neural networks. The pros and cons of these two categories of networks are discussed in details.įor a more detailed outline of this blog post, please read the following introduction section. Next, we divide lidar 3d object detection networks into two categories of networks with input-wise permutation invariance which demonstrate symmetry property to directly process raw point clouds, and networks with point cloud grid representations that rely on ordered structured representations of point clouds. Then, we formally define 3d object detection task and present the common regression and classification loss used to measure the performance of models tackling 3d object detection task. In this blog post, first, we review the data format of lidar point clouds represented in the KITTI dataset. This blog post is best suited for those who have basic familiarity with image-based 2d object detection networks and are interested in learning how the standard methods used in 2d object detection networks are used and tailored towards point cloud 3d object detection task.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |