Please enable JavaScript.

Coggle requires JavaScript to display documents.

Object Detection (:green_apple:IoU (:mega:Intersection over Union, …

- - - - :three:Detection
        
        :checkered_flag:detect up to several objects in a picture
        
        :maple_leaf:YOLO, R-CNN
      - :one:Image Classification
        
        :maple_leaf:traditional CNN
        
        :checkered_flag:classifies a picture
      - :two:Classification w. localization
        
        :maple_leaf:Simplified YOLO, R-CNN
        
        :checkered_flag:detects an object in a picture
    - - :one:YOLO, SSD
        
        直接对预测的目标物体进行回归
        
        一刀流,快
      - :two:R-CNN
        
        1、生成可能区域（Region Proposal） & CNN 提取特征
        2、放入分类器分类并修正位置
        
        二刀流,准
  - - - :cherry_blossom:algorithm
        
        :two:通过卷积得到特征
        
        :three:将特征给SVM和Bbox reg分别做分类与回归
        
        :one:通过selective search为图片找到2k多图片
      - :sweat_drops:problems
        
        :one:每个候选框都要有单独的conv等网络
        
        :zap:不同候选框大小不同,所以必须不同卷积网络
        
        :two:SVM太慢了
        
        :three:training is slow(84h)
        
        :four:47s / image with VGG16
      - :classical_building:structure
    - - :cherry_blossom:将不同大小的输出做spatial pyramid pooling,最后统一成一个fixed-length的向量,然后交给全连接层
      - :classical_building:structure
    - - :cherry_blossom:algorithm
        
        :one:对原图进行卷积
        
        :two:在卷积后的特征图中找到原图的候选框
        
        :three:使用Rol Pooling把输出的候选框弄成相同大小的形状,类似于SPP layer
        
        :four:经过统一的fc层做分类和回归
      - :sweat_drops:problems
        
        :one:RoI产生还是使用传统的方法,产生方法也slow(Fast R-CNN使用selective search只能在CPU计算,且占用2s,即大部分的时间)
      - :classical_building:structure
      - :pineapple:RoI Pooling
        
        :arrow_left:inputs
        
        :one:a fixed-size feature map
        
        :two:an N*5 matrix
        
        :explode:a list of region of interests
        
        :moneybag:<class_id, xmin, ymin, xmax, ymax>
        
        :warning:the regions are in the feature map and their sizes are various
        
        :arrow_right:outputs
        
        :explode:a fixed size image
        
        :icecream:example
        
        :link:Region of interest pooling explained
        
        :cherry_blossom:algorithm
        
        :explode:对于特征图里的每个region,要获得m*n的输出,平均/向下取整对行和列分成m份或者n份,然后取最大值
      - :question:questions
        
        :one:Which step should the selective search be applied to? Before the conv or after?
    - - :star:Fast R-CNN and Faster R-CNN
      - :one:speed comparison
- - - - :snowflake:methods
        
        :one:sliding window
        
        :sweat_drops:large inputs
        
        :two:region proposal
        
        :icecream:selective search
        
        :link:reference
        
        :cherry_blossom:algorithm
        
        :explode:先初始化regions,两两计算相似度,一直取最高的然后合并与剔除旧的,最终留下的就是region了
        
        :sweat_drops:只能用在CPU上,并且需要整整2s
        
        :checkered_flag:为sliding window产生候选框,加快速度
        
        :sweat_drops:本身也很慢
        
        :three:regression
        
        :sweat_drops:不适用于多物体检测
      - 主要使用图像分割与区域生长(合并)技术
    - - :cherry_blossom:details