Lecture 1
Hubel & Wiesel 1959
found out that
visual processing starts with simple structure( ex. oriented edges) and brain builds up the complexity of the visual information
until it recognizes the complex visual world.
Block world, Larry Roberts 1963
attempt to simplify visual world using simple geometric shapes to recognize them and reconstruct the shape
The summer vision project 1966
attempt to use summer workers in construction of a significant part of a visual system
Vison, David Marr 1970
similar to Hubel and Wiesel study!
Generalized Cylinder, Brooks& Binford 1979
Pictorial Structure, Fischler and Elschlarger 1973
Every object is compsed of simple geometric primitives!!
reduce complex structure into a collection of simpler shapes
Davide Lowe 1987
just toy example levels in 60s,70s,80s
Next Level!
so in 90s, if object recognition is too hard, what about object segmentation first?
Normalized Cut Shi& Malik 1997
using graph theory algorithm
Face Detection Viola & Jones 2001
using Adaboost, real time face detectiion
SIFT & Object REcognition David Lowe 1991
feature based object recognition!
match some parts of a objects(features that are tend to remain diagnostic, invariant) to similar object! => easier than matching whole object
Spatial Pytamid Matching, Lazebnik Schmid & Ponce 2006
start to recognize holistic scenes
from different parts,resolution of the image, take features and put them in a feature descriptor then svm to understand image
Histogram of Gradients (HoG) Dallal & Triggs 2005
using upper idea into human
NEXT STEP
comming to 20s century, images are getting better == better data!
began to have bench mark datasets
Pascal visual object challenge
20 object classes benchmark dataset
IMAGENET
most of ml models are very likely to overfit.
began for two reason
1. to recognize world of all the objects (many objects)
2. to overcome the ml bottleneck of overfitting
took about 3 years
CHALLENGE BEGINS
success if top 5 results contain the correct label
error rate steadly decreases!
drop in 2012 is significant!! using CONVOLUTIONAL NEURAL NETWORK!
Alexnet == super vison
LeCun 1998
compuations developments, gpus => larger architectures, models
large high quality data!
made CONVOLUTIONAL NEURAL NETWORK WORK WELL!
Reference
