visual computing theory and engineering...

39
Visual Computing Theory and Engineering Applications Li Song(宋利) Bingbing Ni (倪兵兵) http://medialab.sjtu.edu.cn Shanghai Jiao tong University Welcome all of you!

Upload: others

Post on 25-Jul-2020

0 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Visual Computing Theory and

Engineering Applications

Li Song(宋利)

Bingbing Ni (倪兵兵)

http://medialab.sjtu.edu.cn

Shanghai Jiao tong University

Welcome all of you!

Page 2: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

What is visual computing?

.

Design algorithm or system that can interpret or

processing images/video

1. Interpret: Patter

Recognition, Image Search,

Video Mining; Data ->

Semantic

2. Processing:

Filtering/Transform/Analysis;

Signal -> Signal or Feature

3. “Visual” and “Computing” :

Real problem need both

Page 3: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Why study visual computing?

.

Images and movies are everywhere

Our 80% information come from them

Fast-growing collection of useful applications

3D Contents Creation : Avatar

Smart Surveillance : Object Video

Movie and TV post processing : IMAX

Greater understanding of human vision

how does object recognition work?

Page 4: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Let us see some

interesting images firstly!

Page 5: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Parallel Lines?

Page 6: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Crossing Lines?

Page 7: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Circle?

Page 8: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Which bigger?

Page 9: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Which longer?

Page 10: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Static or moving picture?

Page 11: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Do you see waving?

Page 12: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

What do you see?

Page 13: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

What do you see?

Page 14: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Impossible things?

Page 15: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

What are related branches?

There is still not a general theory for vision, problem driven solution now…

Page 16: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

What this course is about?

3D Reconstruction View Marr’s vision theory

Edge extraction ( and segmentation )

Geometry( and Motion, Depth)

Patter Recognition View Invariance

Detection

Tracking

Recognition

Computational Neuroscience View Human vision system

Neural Computing

Page 17: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Two free reference textbooks

• R. Szeliski, "Computer

Vision: Algorithms and

Applications", 2010.

• S. Prince, "Computer vision:

models, learning and inference",

2012.

Page 18: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Marr’s Vision Theory

Page 19: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

3D Reconstruction

Stereo reconstruction 3D from multiple views

Page 20: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

It can only handle simple objects in real world

Edge detection itself is difficult without

information of high layer

The function of Human Vision is more than 3D

reconstruction.

It is found that visual neurons work in the ways of

both bottom-up and top-down.

Challenge of Reconstruction

Page 21: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Detection

Page 22: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Tracking

CAMSGPF PF

Page 23: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Recognition

Search by example

Page 24: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Recognition

Video Index and retrieval

Page 25: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Recognition

Face detection

Page 26: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Recognition

Event detection

Page 27: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

It must be problem dependent

no common method

Prior knowledge is definitely needed:

Related Context: digging multiple cues

Need many samples: learning from samples

Useful low level and mid level will be helpful:

Get important feature will make things easy

Challenge of PR Viewpoint

Page 28: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Neuroscience: final solution?

Page 29: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Deep Learning-A light of hope?

Page 30: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

What is course objective?

Introducing the State-Of-The-Art research

For most of us, build your work on the

shoulders of giants/genius

Making the students capable of implementing the

algorithms introduced by newly published papers.

The skill of writing scientific papers(Publications),

and presenting your work (PPT)

The skill and capability to do research as a team

Page 31: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

How to score your performance?

Show up at class (30%)

Sign each time

Paper Reading Report (30%)

PPT

Final project: (40%)

Review paper from selected topic

Page 32: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Tips: how to learn?

Technology

Engineering

Methodology

Concept & Model

Algorithm & Framework

System & Application

Science

Philosophy

Page 33: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Tips

Teaching materials will include

PPT, and recommend papers

Videos/PPTs from internet

Highly recommendation

Reading more after class, what you get from

class if just beginning

Writing report when reading

Programming is key for your success

Just because we are engineers…

Page 34: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Tips

How to treat Publication?

Very important for: summer interns, job

hunting, get well known, etc.

But!

“An article about computational

science in a scientific publication is

not the scholarship itself, it is merely

advertising of the scholarship. The

actual scholarship is the complete

software environment and the

complete set of instructions which

generated the figures.”

Page 35: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Tips

• Resources

▪ Journals:TPAMI, IJCV; NC, IJML; TIP, TMM; PR…

▪ Conf.:ICCV, CVPR, ECCV; NIPS, ICML; ACCV,

MM, ICIP, ICPR, CVGIP; SIGGRAPH...

• Homepages ▪ Google search:computer vision

• Computer vision homepage(CMU) • Cvonline (online turtorial) • Computer vision Research Groups

(CMU/MIT/BERKELY/Standford/UIUC…)

• Big guys ▪ Just search according to their name and affiliate

Page 36: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Tips

Open Courses/Lectures -http://www.coursera.org/course/

-http://videolectures.net

-…

Page 37: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Tips

Ongoing courses http://vision.stanford.edu/teaching/cs231n/index.html

Page 38: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

Tips

Programming Open sources

Collect…

Assemble…

Integrate…

Page 39: Visual Computing Theory and Engineering Applicationsmedialab.sjtu.edu.cn/teaching/CV/Lec/lec1.pdf · Edge detection itself is difficult without information of high layer The function

More information cab be found at:

http://medialab.sjtu.edu.cn/teaching/CV/CV.htm