CS231n

2) Image Classfication pipline

이미지 분류 - computer vision의 핵심

이미지 classfication이 가능해야 => detection, Segmentation, Image captioning 가능

이미지는 숫자로 구성되어 있다

3001003(height width colorchannel)

Classfier challenge:

시각에 따른 변화/조명/형태의변형/은폐은닉(occlusion)/background clutter/Intraclass variation

Classfier 기본함수

def predict(image):

​ return class_label

이미지 예측 -> 이미지 label

숫자를 분류하는 것과 다르게 이미지 구분에는 명백한 알고리즘이 없다

data-driven approach

image and label - train image classfier - test image classifer

First classfier - Nearest Neigbor Classfier

Remember all training images and their labels in Memory

Predict the label of the most similar trainging image by metric distance

Example -CIFAR 10

10 labels

50000 trainging images

10000 test images

How to Compare?

1. L1(Manhattan) distance
2. L2(Euclidean) distance

NN Classfier linearly depend on the size of the train data

-> ANN,FLANN 이용해 속도 높이기

-Hyperparameter -> 실험을 통해서 알아보기

• what is the best distance to use?

• what is the best value of k to use?

  => Must try alll

train -> test : bad idea, Test set use only at the end

train -> validation data(from train data) -> test data :

=> use Cross validation

k-NN never used

terrible performace at test time

distance metric are not very unintutive

Linear Classfication

• Parametirc approach
  • f(x,W) x : image/W : parameters
  • f(x,W)(10 * 1) = W(10 * 3072)x(3072 * 1)+b(10 * 1)