[ CS231n ] Introduction to Convloutional Neural Networks for Visual Recognition

Visual data on internet 

- really hard to understand

Youtube

- 3초 동안 15시간 분량의 비디오가 유튜브에 업로드됨

 

This lecture focuses on neural networks and CNN and visual applications

 

TOPIC

A brie history of computer vision

Evolution's Big Bang

- fossil studies : 10 million yeasr - the number of animal species 🔝

-around 540 million years : animals began to have eyes - they got VISION ❗️ =the biggest sensory system 

History of Human's vision:

1600s : camera obscura 

Hubel and Wiesel (1950~1960s)'s work

- used electrophysiology

- "what was the visual process mechanism like in mammals, primates...?"

- 고양이 뇌로 실험:

   visual process starts with simple structure -> the brain builds up the pathway => understand complex visual world

Block World by Larry Roberts,1963

The Summer Vision Project, 1966

David Marr,1970s

Generalized Cylinder, Brooks & Binford, 1979 && Pictorial Structure, Fischler and Elschlager, 1973

   Every object is composed of simple geomertric primitives

IF it is hard to recognizae object maybe we should first do object segmentation

HOW? 

1. Take an image

2. Grop the pixels into meaningful areas //can't tell if the pixels were a person or a background

 

"AdaBoost Algorithm"

SIFT, David Lowe, 1999

    - find critical features and match the features to a similar object

Spatial Pytamid Matching

The quality of image improved meanwhile

PASCAL Visual Object Challenge: 20 object recognition

  => Are we ready to recognized things in the real world?

  =>AdaBoost: high possibility of overfit !

The Image Classification Challenge:

    Result = Top 5 possible object list- if answer is in the list? CORRECT

    2012: error rate dropped! Winning algorithm =CNN(Convolutional Neural Network)

2012 had a remarkable breakthrough in neural network but it was not invented in 2012.

There were foundational works: 1998 LeCun et al. 2012 Krizhevsky et al. //had architerctural similarities

Factors contirbuted to advancement:

  1. Computation -GPU 2. Improved data

The goal of computer vision:

   If people see a picture even for a very short amount time , they can tell or even make stories upon the image

   People can not only sort the objects in the image, but they see and understand the situation.