深度学习入门(4) -Object Detection 目标检测

Object Detection

Output:

1. category label from fixed, known set of categories
2. bounding box (x, y, width, height)

If only one object is needed to be detected -> add FC layer to the Net pretrianed on ImageNet

Sliding Window

apply a CNN to many different crops of the image, CNN classifies each crop as object / backgroud

but too many windows!! and may detect repeatedly

we need region proposals to find a small set of boxes that are likely to cover all the objects

“Selective Search” quick to generate 2000 regions

R-CNN : Region-Based CNN

1. Region proposals
2. warped the image to fixed size 224*224
3. forward each region through ConvNet independently
4. output a classification score and also a Bbox of 4 numbers, using the following algorithm
   

Measurement of boxes (IoU)

$IoU=\frac{\text{Area of Intersection}}{\text{Area of Union}}$