3D point cloud course (VI) -- 3D target detection

1.  Image target detection

Box representation ： LWH 、 Center point 、 toward 、 Category

Ideas ： Detect first and then classify

 1.1 Evaluate the quality of the test

 Precision： Probability of detecting the right       Recall： Probability of detection     It's hard to have both

• AP（Average Precision）  mAP(Mean over all categories)

 

• NMS

Each selector is followed by NMS, Select the best detection box

Select the box with the highest confidence , Then calculate the overlapping range of other boxes and this box IOU, If it is greater than a certain threshold, it will be rounded off .

 1.2 Methods of object detection

1.2.1 Two-Stage

First locate and then detect

• RCNN ： In essence, it is a clustering frame , Then classify the boxes . It is slow to classify each box

•  Fast RCNN

First, extract features from the image , Again ROI Pooling Zoom to uniform size , after MLP. Classification neural network is working in Feature map Upper .

ROI Pooling The specific process is as follows , No matter what the original size of the box is , Finally, they are scaled to 7*7.

ROI Pooling There is a loss of precision , Put forward ROI Align linear interpolation

• Faster RCNN 

The box is given by deep learning

• Mask RCNN  Instance segmentation

 1.2.2 One-Stage

  Think of the background as a class , Out of every place 3 individual anchor box , Classify each , Calculation location . No, first judge whether there is an object in the box .

 

2. Point cloud target detection  

Projection angle

Three dimensional grid

Projection

2.1 VoxelNet

Three dimensional convolution pointnet++

There are many squares without dots , No afferent neural network is needed  , The acceleration method is as follows

• Data Augmentation

1、 Rotate the whole point cloud , But don't rotate too much

2、 Translate and rotate the target , Not too much  

2.2 PointPillars

A pile of columns in the space , Each column is compressed into a plane  

• Focal Loss

Solving the problem of unbalanced categories , Normally, less categories will be ignored , This method turns attention to small categories .

 

2.3 Point-wise operation

2.3.1 PointRCNN

1. Get the of each point feature vector
2. Divide the foreground and background , Make one for each prospect proposal
3. every last box do ROI Pooling

All operations are input in the form of points  

2.4 Vison Fusion

2.4.1 Frustum PointNet

Project the frame on the image onto the point cloud , Vision and LIDAR It is difficult to align in time and space .

2.4.2 PointPainting