Voxel based and second network learning

VoxelNet

• principle ：

  Divide the 3D point cloud into a certain number of Voxel, Then random sampling and normalization , Non empty Voxel Feature extraction results in Voxel-wise Feature,3D Abstract features of convolution middle layer network , Last use RPN Object classification detection and position regression
  

• Feature Learning Network

1. Voxel partition
2. grouping
3. Random sampling






Convolutional Middle Layers
This layer has a lot of calculation , It is also the biggest problem of the whole network ,Second Replace this network with sparse 3D Convolution network , The computing speed is greatly improved

Region Proposal Network
Output probability score diagram and regression diagram

SECOND: Sparsely Embedded Convolutional Detection

SECOND Three contributions of ：

1. take VoxelNet Three dimensional intermediate convolution layer is replaced by sparse convolution layer , Greatly improve the computing speed
2. Added a direction regression , Improve the convergence speed of model training
3. Data to enhance , Splice some truth point cloud data

1. Sparse convolution reduces the amount of computation , But there will be sub epidemic inflation , Sub popular convolution can guarantee the original resolution , But it can not improve the receptive field

RPN Directional regression

Data to enhance