Point cloud perception algorithm interview knowledge points (I)

1. In a nutshell pointrcnn Based on the bin The return of the
answer ：pointrcnn Is the first one based on pure point cloud anchor free Two stage approach , stay pointrcnn Previous papers used anchor base, It uses IOU Calculate whether it belongs to the positive sample . then pointrcnn in consideration of , For example, the distance between the point at the front of the car and the point at the door and the center is still very large , Direct regression is not good . Because classification is better than regression , therefore pointrcnn Just through bin The object center regression problem is transformed into a classification problem , The center of the object falls there bin Inside , And then we can go back to the central coordinates here bin The offset inside , So we can get the central coordinates of the object . In this way, the classification and regression method is more precise than the direct regression method , better .

2. Transformer Three vectors in Q、K、V Calculation method

answer ： In the coding module , The self attention mechanism of the encoder first converts the input vector into three different vectors , Query vector 、 Key vector 、 Value vector , Then the vectors from different inputs are packed into three different matrices , Then the attention function between different input vectors is calculated by the following four steps ：

The first step is to determine the degree of attention paid to other location data when encoding the current location data , Calculate the fraction between two different vectors S. The second step is to score S Standardize to Sn, Make it have a more stable gradient , In order to better train , The third step will use softmax The function divides the score Sn Convert to probability P. The fourth step is to obtain the weight matrix , Each value vector V Multiply by the probability P, A vector with a high probability will be paid more attention by subsequent layers . This process can be uniformly expressed as ：

3. In the convolution layer BatcnNorm The role of

answer ： Increasing the learning rate makes learning faster ; Less weight dependent 、 The initial value of the offset ; It can restrain the occurrence of over fitting phenomenon .
notes ：BatchNorm, The average value of the input data is 0、 The variance of 1（ Appropriate distribution ） Regularization of , Then we translate and rotate the regularized data .

4. How to solve the problem of imbalance between positive and negative samples , How to calculate the solution ？

answer ： Usually use focal loss Loss function to solve the imbalance problem of positive and negative samples . The loss function reduces the weight of a large number of simple negative samples in training , It can also be understood as a kind of difficult sample mining .
Focal loss It is based on the cross entropy loss function , First, the binary cross entropy loss function is ：
y/ Is the output of the activation function , So in 0-1 Between . It can be seen that the common cross entropy loss function for positive samples , The greater the output probability, the smaller the loss . For negative samples , The smaller the output probability, the smaller the loss . At this time, the loss function is slow in the iterative process of a large number of simple samples and may not be optimized . therefore Focal loss Improved binary classification cross entropy ：

The first is introduction gama factor , When gama=0 when , Is the binary cross entropy function , When gama>0 when , It will reduce the loss of easily classified samples , To focus more on the difficult 、 Misclassified samples .

Then add the balance factor alpha, It is used to balance the uneven proportion of positive and negative samples . But if you just add alpha Although it can balance the importance of positive and negative samples , But it can't solve the problem of simple and difficult samples .

5. KITTI Data set easy、moderate、hard According to what definition ？ P-R What is the curve , How to calculate ？ AP_R40 What is it? ？ threshold 0.7 What is it? ？ How to calculate ？

answer ： KITTI Data set easy、moderate、hard Depending on whether the dimension box is occluded 、 The degree of occlusion and the height of the box are defined , The specific data are as follows ：
Simple ：
Minimum bounding box height ：40 Pixels , Maximum occlusion level ： Completely visible , Maximum truncation ：15%
secondary ：
Minimum bounding box height ：25 Pixels , Maximum occlusion level ： Partial occlusion , Maximum truncation ：30%
difficult ：
Minimum bounding box height ：25 Pixels , Maximum occlusion level ： It's hard to see , Maximum truncation ：50%

P-R In the curve P It stands for precision（ Accuracy ）,R It stands for recall（ Recall rate ）, It represents the relationship between accuracy and recall , In general , take recall Set to abscissa ,precision Set to ordinate . Every one of them P-R Each curve corresponds to a threshold （ Such as threshold 0.7 etc. , It's through IOU Calculated ： The intersection of two boxes divided by the Union ）
computing method ：

AP_R40

KITTI The official ranking list used AP_R11, namely AP_R11 = {0,0.1,…,1}. However , Even if the tangents match , contain 0 It will also lead to an increase in average accuracy of about 9％. To avoid this apparent performance improvement ,kitti The government has adopted a new 40 spot AP（AP_R40） To revise metrics and leaderboards , Exclude “0” And quadruple dense interpolation prediction , To achieve better results Precision / Recall Approximate value of the area under the curve . Now more and more papers begin to use AP_R40 To compare the performance of the algorithm model .