Object Detection
Object Detection
Four steps for Computer Vision Tasks :
Classification , Classification + Localization , Object Detection , Instance Segmentation
one-stage
two-stage
bbox : Bounding box 包含物体的最小矩形 ( x , y , w , h )
x , y 默认为左上角,可自定义
w , h 宽和高
IoU : Intersection over Union ---- Intersection / Union
NMS : Non-Maximum Suppression 非极大值抑制( search for local maximum )
sliding windows :
Classification && Localization
R-CNN ( Region Based CNNs )
-
Input images
-
Extract region proposals ( ~2k ) 提取 Selective Search Algorithm
-
compute CNN features
-
classify regions ( SVM )
Selective Search Algorithm ( Based on Greedy Principle )
combine smaller segmented regions to generate region proposal 分割区域
input : an image
output : generate region proposals
advantage over random proposal generation 产生 :
process :
-
generate initial sub-segmentation of input image.
-
Combine the most similar two ( What would happen if two is changed into other number ? ) bounding boxes into larger ones recursively 递归地 , until there is only one image left .
color similarity , texture similarity , region size , etc.
-
Use these larger boxes to generate region proposals for object detection.
CNN
model ----- AlexNet model
stride of 4 四步 :
max pooling :
input : ( 227 , 227 , 3 )
softmax layer 层 :
SVM ( Support Vector Machine ) :
bounding box regressor 回归量 :