algorithm framework: mtcnn+sphereface+ncnn
Recently, I was doing face recognition. I saw an open source project, sphereface, on github. The approach is to detect the five coordinates of the face frame, eyes, nose and corners of the mouth through mtcnn. According to the five coordinates, I realize face alignment, pick out the aligned face, and pass it to sphereface. Finally, I extract the feature of the face through sphereface. Through comparing the feature differences, I get whether the faces match.
KNN classification algorithm and decision tree classification algorithm in the first two chapters are the final classification results that predict the instance, but sometimes the classifier will produce wrong results;The Naive Bayesian classification algorithm to be studied in this chapter is to give an optimal guess result and also give a probability estimate of guess.
1 Preparation Knowledge: Conditional Probability Formula
I believe that students who have studied probability theory will never be unfamiliar with probability theory.
reprinted from http://blog.csdn.net/jasonding1354/article/details/50562513.
Content Summary Disadvantages of Training Set/Test Set Segmentation for Model Verification K Fold Cross Verification How to Overcome Previous DeficienciesHow can cross-validation be used to select adjustment parameters, select models, and select featuresImproved Cross Validation
autoaddition is a single-purpose operator divided into: pre ++(a++), and post ++(++a), which is usually used in assignment statements.
a++: assign a value before performing a=a+1 ++a: perform a=a+1 before assigning Example 1:
int a=1; int b=a++; a=1; int c=++a; System.out.println("b="+b+",c="+c); Results: b=1,c=2
int i=1; i=i++; int j=1; j=++j; System.out.println("i="+i+",j="+j); Results: i=1,j=2
Analysis: i=i++ Store the old value of I in temp and then let I add it by itself.
1. Problems Solved in the Paper 2. What is Active Vision 3. Differences from visual saliency, visual attention (visual saliency and attention) and prediction related features (equivalent to differences between this article and previous methods) 4. Solution of the thesis 5. To what extent has the solution in the paper solved this problem? 6. Other Issues Not Considered 7. Experiments
Training errors and actual errors were introduced in the previous blog post.When the training errors are very low, but the actual errors are very high, it shows that the classifier we constructed has over-fitting.The reason for over-fitting is that the classifier we designed is too complex to record all the classification data.This leads to poor scalability of the classification model.Therefore, the complexity of classifier is closely related to the scale of training data.
Hong Ruan recently opened the SDK engine for face recognition (free of charge), which happened to have Android version and experienced a wave.Let's talk about Android SDK usage experience: ArcFace Hongruan Face Recognition Engine The currently open versions of human face comparison (1:1) and face retrieval (1: N) can be selected according to the application scenario. Face Retrieval is divided into small network (within 100 persons), medium network (within 1000
Recently, I found that I logged into the background management system and found that the authentication video uploaded was too vague. As a result, I found that I did not set the frame frequency for MediaRecorder. // 设置帧频率，录制视频会更加清晰 mRecorder.setVideoEncodingBitRate(5*1024*1024); 1. Open the camera directly to get the desired bitmap. About why YUV Image: https://blog.csdn.net/illidantao/article/details/51366047 is used