PointNet网络结构详细解析
一、重要知识点
- Transforming point clouds data to regular 3D voxel grids or collections of images, however, renders(cause to be) data unnecessarily voluminous(length, vast) and introducing quantization artifacts, obscure(conceal) natural invariances of the data.
- PointNet learns to summarize an input point cloud by a sparse set of key points, which roughly corresponds to the skeleton of objects.
- 将点云体素化会改变点云数据的原始特征,造成不必要的数据损失,并且额外增加了工作量,而 PointNet 采用了原始点云的输入方式,最大限度地保留了点云的空间特征,并在最终的测试中取得了很好的效果。
- A symmetric function is invariant to the input order. For example, + and * operators are symmetric binary function.
- Treat the input as a sequence to train an RNN.
- 相同的点云在空间中经过一定的刚性变化(旋转或平移),坐标发生变化,但希望网络都能正确的识别出物体(Special Transform Network, STN),但最终实验结果和后续论文PointNet++表示,STN并无多大作用。
- 基本思想:对输