Image Classification pipeline & Liniear classification

Image Classification : A core task in Computer Vision

• Semantic Gap : represent image to numbers

• Challenges

  • Viewpoint Variation( = The numbers are changed when the camera moves )
  • Illumination
  • Deformation( = different pose & position )
  • Occlusion (in medicine, something that blocks a tube or opening in the body, or when something is blocked or closed)
  • Background Clutter ( = look similar to background )
  • Intraclass variation

• An image classifier (API in python)

  • no obvious way the algorithm for recognizing a cat, or other classes.

def classify_image(image):
	# Write magical Some codes
	return class_label

Data-Driven Approach (way to make algorithm work)

1. Collect a dataset of images and labels
2. Use Machie Learning to train a classifier
3. Evaluate the classifier on new images

First classifier : Nearest Neighbor

import numpy as np
class NearestNeighbor :
	def __init__(self):
    	pass
    # Memorize training data
    def train(self, X, y): 
    	self.Xtr = X
        self.Ytr = y
    def predict(self, X):
    	num_test = X.shape[0]
        Ypred = np.zeros(num_test, dtype = self.ytr.dtype) #output data type matches input type
        # For each test image: Find closest train image, Predict label of nearest image
        for i in xrange(num_test):
        	distances = np.sum(np.abs(self.Xtr - X[i,:]), axis = 1) #using the L1 distance (sum of absolute value differences)
            min_index = np.argmin(distances) #get index with smallest distance
            Ypred[i] = self.ytr[min_index]
        return Ypred

• in this classifier Train process is faster than Predict process (because Train Time complexity function is O(1), Predict is O(N))
  • This is bad, we want faster classifier in prediction, slow for training is ok.
• Distance Metric to compare images (method to compare train and test data)
  • L1 distance : easy idea to compare images ( just compare each fixel values )

K-Nearest Neighbors

• Instead of copying label from nearest neighbor, take majority vote from K closest points
  
  can see that the larger K, the more smooth decision boundary
• K-Nearest NEighbors : Distance Metric
  • L1(Manhattan) distance vs L2(Euclidean) distance (has more natural boundary)

• L1 tends to follow coordinate axis because L1 depends on our choice in coordinate system.
• L2 doesn't care about coordinate axis. It just put decision boundary naturally.

Hyperparameters

What is the best value of k and distance to use?
We call this k and distance hyperparameters.

• Must try them all out and see what works best.

• Setting Hyperparameters

  • Use all data to train data > K=1 always works perfectly on training data.
  • Split data into train and test and choose hyperparameters based on result of test data. > also not good. No idea to perform on new data.
  • Split data into train, test and validation data. and choose hyperparameters on validation and evaluate on test. > great idea!
  • Cross-Validation : Split data into folds, try each fold as validation and average the results. Often used when quantity of data is small.
    

• K-nearest Neighbor is work quite exactly. However, it can not used on images.

  • Very slow at test time & Distance metrics on pixels are not informative.
  • Curse of dimensionality

Summary

• In Image classification we start with a training set of images and labes, and must predict labels on the test set.
• K-Nearest Neighbors : predicts labels based on nearest training examples
• Distance matric(L1,L2) and K(in K-Nearest Neighbors) are hyperparameters
• Choose hyperparameters using the validation set; test set is used only once at the end!

Linear Classification : important to build whole NN

Parametric Approach : Linear Classifier by using weights

more detail will be introduced lecture03