Object Detection Study Notes

研究背景

Three popular directions of computer：计算机视觉、自然语言处理(NLP)以及语音识别

计算机视觉研究方向：

物体识别和检测

Find common objects in pictures,and output its category and location,比如：人脸检测,车辆检测

识别：Classify which samples are the target

检测：Detect informative features from random noisy images

word segmentation

Classify each pixel in the image,It is equivalent to segmenting the target in the picture

运动和跟踪

The position and size of the target are given at the beginning of the video,Then follow it up in the video follow-up

视觉问答

The purpose is based on the input image,由用户进行提问,而算法自动根据提问内容进行回答

The meaning of object detection

From a research perspective,It is one of the fundamental problems of computer vision,It is the basis of many high-level vision tasks

High-level vision tasks：人脸识别、Pedestrian re-identification、目标跟踪、图像分类

应用角度看,A wide range of application requirements have been expressed

Object detection applications：人脸解锁、视频监控、Statistics of people entering and leaving、辅助驾驶、自动驾驶

发展脉络

 

 The single-stage method requires only one correction of the anchor box to get the final result,The multi-stage method requires multiple corrections.

 传统方法

Take advantage of handcrafted features+分类器,以Sliding window method在图像金字塔Iterates over all positions and sizes,进行物体检测

The sliding window method traverses all the positions,Image pyramids are traversed at different sizes

常用数据集 

General object detection dataset

人脸检测数据集

评价指标

检测精度

Divide and overlap ratios（IoU）= 交集面积 / 并集面积

召回率 = The number of ground-truth annotations that are recalled（Number of callout pairs）/ The number of true annotations（Number of actual targets）

精度 = 真正例的数量（Number of callout pairs） / The number of test results（Quantity detected）

The recall is to see whether the detection is comprehensive,Accuracy depends on whether the detection is accurate

平均精度均值（mAP ）  All species average precision（AP）的均值

平均精度 精度-召回率曲线下的面积

漏检率（Miss rate） =  漏检的数量 / The number of true annotations

Average number of false picks per image（FPPI）= The number of missed checks / 图片数

检测速度

前传耗时（ms） = The time taken from inputting an image to outputting the final result

每秒帧数（FPS）= The number of images that can be processed per second

作业

Overfeat：Integrated Recognition, Localization and Detection using Convolutional Networks

RCNN学习笔记(9):OverFeat:Integrated Recognition, Localization and Detection using Convolutional Networks_Focusing on your own Mind :) －> 站在巨人的肩膀上-CSDN博客

R-CNN：

R-CNN论文详解（论文翻译）_Cheese的博客-CSDN博客_r-cnn

fast RCNN:

Fast R-CNN论文详解_WoPawn的博客-CSDN博客_fasterrcnn论文

faster-RCNN

Faster R-CNN文章详细解读_Michael’s Blog-CSDN博客

FPN：feature pyramid networks for object detection

FPN（feature pyramid networks）算法讲解_AI之路-CSDN博客_feature pyramid

YOLO：

YOLO系列算法精讲：从yolov1至yolov4的进阶之路（2万字超全整理,建议收藏！）_不积跬步,无以至千里！-CSDN博客_yolo系列算法

SSD：

SSD原理解读-从入门到精通_QQBrother's column-CSDN博客_ssd原理

R-FCN:object detection via region-based fully convolutional networks

R-FCN算法及Caffe代码详解_AI之路-CSDN博客_fcn代码详解

DCN:Deep & Cross Network for Ad Click Predictions

Recommendation system deep learning articles-DCN网络介绍(2)_来自Daisyand her one-way ticket-CSDN博客_dcn网络

RetinaNet:Focal Loss for Dense Object Detection

RetinaNet论文理解_JustForYou的博客-CSDN博客_retinanet

Mask R-CNN:

【Mask RCNN】论文详解(真的很详细)_Coffee flavored coffee-CSDN博客_mask rcnn

召回率：4 / 5

精度：4 / 6

漏检率：1 / 6