QA Official

[Paper] CVPR 2016, 2017,2018 best paper (with essay download Address) 2019-04-26
1, 2018 Awards Best Paper Award"Taskonomy: Disentangling Task Transfer Learning" by Amir R. Zamir, Alexander Sax, William Shen, Leonidas J. Guibas, Jitendra Malik, and Silvio Savarese.Address of essay download Best Student Paper Award"Total Capture: A 3D Deformation Model for Tracking Faces, Hands, and Bodies" by Hanbyul Joo, Tomas Simon, and Yaser Sheikh.Address of essay download Honorable Mention"Deep Learning of Graph Matching" by Andrei Zanfir, Cristian Sminchisescu.Address of essay download"

[paper notes] face key point detection _ abridged version _2016 2019-04-26
201610091、Bulat A, Tzimiropoulos G. Two-stage Convolutional Part Heatmap Regression for the 1st 3D Face Alignment in the Wild (3DFAW) Challenge[J]. arXiv preprint arXiv:1609.09545, 2016. The algorithm described in this article won the first place in the first 3DFAW competition.3DFAW is a competition for 3D face key point detection. It provides training database for the competition (including face pictures and 3D face key point labeling information). This method is based on the work of Convolutional PartHEAT MAP REGRESSION.

[paper reading notes] batch normalization _ accelerating deep network training by reducing internal co variant shift 2019-04-26
prerequisite knowledge:covariate shift, covariate change, i.e. the change of input variable x itself [the input distribution for a learnable system is changing.Generally speaking, the input of each layer of subnetwork is constantly changing during the training process of a network]Batch-Normalization enables the distribution of network layer inputs to present a standard normal distribution [mean value is 0, variance is 1], which can speed up the training of the network [the distribution of each layer of network is different, the required learning rate lr is different, the network system needs to use the lowest lr to ensure the convergence of the model], and can also partially solve the problem of gradient explosion [by reducing the dependence of gradient on the initial value of parameters].

conditional random fields and Code Implementation 2019-04-26

convert CompCars dataset to. tfrecord format 2019-04-26
Preface: Either object detection or object classification needs to process the original image to generate data suitable for training.Object classification-requires an image and its corresponding label.Object detection-requires an image and its corresponding label and (xmin,ymin)(xmax,ymax)Therefore, the data level usually consists of two partsAnnotation-information of stored images, i.e. file name and label sum (xmin,ymin)(xmax,ymax) corresponding to the file name, are usually in the form of xmlImage-image file Common Data Sets: the

front-end face questions summary pen questions 2019-04-26
Front-end Interview Topic 1.the difference between div and span? div is a block-level label and span is a row-level label 2. What are the values of position in html? The default value is What? values: static, relative, fixed, absolute Default: static 3. Which three layers are the front page composed of, respectively, What?The role is What? Front-end Page Composition: Structure Layer, Presentation Layer and Behavior Layer structural layer Created by

how to read papers-summary of "how to read apaper" 2019-04-26
Abstract Researchers need to spend a lot of time reading papers, not only teachers and researchers, but for some students, especially newly enrolled graduate students, it takes a lot of time to read papers, and often no benefits are seen. This article proposes a three-pass method to help us read an article. It also introduces how to do a literature survey. THE THREE-PASS APPROACH First of all, we are not going to read the paper from beginning to end in detail.

neural network improves mnist recognition rate 2019-04-26
Follow tensorflow's introductory study to build a neural network to improve the mnist recognition rate, and finally to the correct rate close to 1. Refer to The site code for basic reference, type up and understand the process by yourself, and mark the place in Chinese if you don't know it. # -*- coding:gbk -*- import input_data import tensorflow as tf mnist=input_data.read_data_sets("MNIST_data/", one_hot=True) #添加x作为占

summary of machine learning -GBDT, XGBOOST parameters 2019-04-26
Experimental Data Set Selection 1. Classification Data Select load_iris The Irises Data Set: from sklearn.datasets import load_iris data = load_iris()[[10, 25, 50]][[10, 25, 50]] list(data.target_names) list(data.feature_names)2.回归数据选取 from sklearn.datasets import load_boston boston = load_boston() print( boston.feature_names 将数

BN and dropout: Some Problems and Considerations 2019-04-25
1, BN scale Initialization scale is generally initialized to 1.0. It is associated with sqrt(2.0/Nin), where Nin is the number of input nodes, if random positive distribution is used to initialize weights when using relu activation function.That is, it is larger than the normal method by the square root of 2 (because half of the data after relu becomes 0, so it should be multiplied by the root 2).