[Quiz]
Title Description
About K-means clustering algorithm, please answer the following questions:
-
K-means clustering is supervised or unsupervised clustering? (2 minutes)
-
Write N samples (X = (x1, ..., xN)) to optimize the objective function poly K-means clustering algorithm K class. (6 points)
-
Please write pseudocode clustering process. (8 points)
-
Wherein the sample dimensions is assumed as D, describe Kmeans time complexity of algorithm. (4 points)
Answer:
1, unsupervised
2, the objective function: F (X, K) = i = 1Kj = 1Ni (xj-μi) 2
3, the clustering process:
初始化：从N个样本中随机选择K个作为初始聚类中心；
For t=1:T（此处，T为最大迭代次数）
将N个样本按距离最近原则分配给K个聚类中心；
迭代更新聚类中心；
如果达到终止条件，如全部样本归类无变化，或者样本点到聚类中心的平均距离变化率较低，则退出
4, the time complexity: TNKD wherein T is the iteration number, N is the number of samples, K is the number of cluster center, D is the dimension of the sample
[Quiz]
Title Description
What camera model parameters, write the three-dimensional coordinates of the point to the image projection formula. Which has several lens distortion coefficient corresponding distortion correction formula expression?
Answer:
within the parameters of the camera model is important are: the main point, focal length, distortion factor. (3 points)
Point two-dimensional three-dimensional image projected formula: (10 minutes)
Lens distortion: radial distortion and tangential distortion. (2 minutes)
Expression of radial distortion correction formula: (3 min)
Expression of the tangential distortion correction formula: (2 minutes)
[Quiz]
Title Description
System question (20 minutes)
In the field of automatic driving, the spatial position of the obstacle is very important information. Without the ability to have the radar range finding sensors, how to calculate the monocular camera only by the distance of the target on the ground?
Answer:
using IPM (Inverse Perspective Mapping) method to achieve, IPM projection as follows:
the original image
an image after the conversion IPM
There are two main ways to generate image IPM
Scheme One: the ground is assumed flat, direct calculation of the homography matrix H is projected manner. 4 groups selected automatically or manually marked points and the points corresponding to the three-dimensional space, the original image to calculate the IPM converted image homography H. Thus each of the ground plane on the image coordinate system that can be calculated by multiplying the real three-dimensional space coordinates after the H matrix.
Option II: the same assumptions need to be flat ground, by calibrating the camera intrinsic parameters and extrinsic parameters, in terms of a relationship between the image projection plane and the three-dimensional space. Formula derivation points can be given.
[Quiz]
Title Description
Current methods for detection by the convolutional neural network can be divided into one-stage and two-stage, corresponding respectively to write algorithms to understand. What are the differences in the two types of detection algorithms in common?
Answer:
One-Stage: yolov1, yolov2, yolov3, the SSD, RetinaNet (2 minutes)
Two-stage：Fast R-CNN、Faster R-CNN（2分）
Two-stage common detection algorithm to faster r-cnn example, the use of a complex network for classification and regression of each candidate region; after a greater number of feature channels ROI pooling, resulting in memory consumption and computation are relatively Big.
Common One-stage detection algorithm, from a structural point of view the network just rpn Networks Combined, the equivalent of the first phase of faster rcnn, so the one-stage main advantage is speed. The result is a prediction and classification of the target position from the feature map regression, and some also adopted the concept of anchor. After the above-described two-stage results roi pooling further refinement, so the detection accuracy of two-stage method are generally relatively high. Another view is that, RPN portion of the two-stage balanced positive and negative samples equivalent of doing this is two-stage detection of a relatively good reason. one-stage algorithm for small target detection less effective, if all of the anchor are not covered by this goal, then the target will be missed. If a large anchor coverage that goal, then the larger receptive field will weaken the real characteristics of the target, the score will not be high. roi pooling two-stage target will do a resize algorithm, wherein a small target is amplified, wherein the profile is also more clearly, and therefore is also more accurate.
Above the red mark, a little to answer four points, if the answer is reasonable, is not the answer to the above is also a plus.