🍒Lecture 12-13

● Object detection

○ IoU

Intersection over Union

Decent尚好的

○ Multi-task loss

○ Sliding window, Regions proposals

Fix: Regions proposals

● RCNN (Region-based CNN)

Region of interest-> warp扭曲

what

Try to do？

How to make it fast

○ Region classification, box regression

Given a image-》predict a transfornation of the box

Negative part 不会被train

（background

//

● NMS（Non-Max Suppression）

What it rty to meaure

（Non-Max Suppression）

用于solve overlapping detections

● mAP （mean Avarage Precison）

What is

（mean Avarage Precison）

AP of cat

AP of dog

求平均

//

* Precison： of all the detection you have ， how many are correct

Sofar only one detection（it‘s correct）

• Recall： of all the ground truth，how many we found

perfectAP：percison1at every recall point

Eg自动驾驶，recall更重要！（find all people

● Fast-RCNN

先进入convNet 在crop

○ ROI pooling/align

Roi polling is

//不snap，直接divide；；；sample some points；双线性插值（bilinear interpolation

//roi align在 maxpooling