lecture2 preview, keyword

수업 후에 프리뷰를 적어 이상하지만, 키워드 일단 적어보자

 

lecture2. Image Classification Pipeline

python / numpy tutorial

 

1.Image Classification

 

sementic gap : human vision vs digital image data

challenges

 - viewpoint variation

 - illumination

 - deformation

 - occulusion

 - background clutter

 - interclass variation

An image classifer(Recognition)

 - no obvious way of hard-code

 - edge/corner

 - Data Driven Approach

 

1st classifier : K-NN, Nearest Neighbor(memorize / predict)

 - dataset : CIFAR10

 - Distance Metric to compare image : L1/L2 distance metric

 - L1 distance

 - time complexity O(1), O(N)

 - hyperparameter : best k/distance choice, problem dependent

 - dataset : train/validation/test

 - k-fold validation

 - knn - image never used(curse of dimensionality)

 

Linear classification

 - parametric approach

 - hard case

 - defined score function