6.819 / 6.869: Advances in Computer Vision
Website: h=p://6.869.csail.mit.edu/fa15/
Aude Oliva
Lecture TR 9:30AM – 11:00AM (Room 34-‐101)
K-0786.7380&L&A)0&
)6*/&M>1+,&N81-.15,>&B/0/,8.<&
O.1/-P07&
3>1+,Q417"/*6&
MR./&<3680(&CSITHCS&
A6/0*,U&$$I$S&54&38&&
;U&,5531-74/-7&A'A<&
V606?&)U7,8&N307*3.738,>&B/0/,8.</8&
U606?Q.0,1>"417"/*6&
MR./&<3680(&B334&AWT&
A6/0*,U&@I!&54&
)*17U,&X<30>,&YY2O&Z8,*6,7/&076*/-7&
[<30>,Q.0,1>"&417"/*6&
MR./&<3680(&B334&AWT&
E3-*,U&@I!&54&
W3>/1&\<36&YY2O&Z8,*6,7/&
076*/-7&
;3>/1Q417"/*6&
MR./&<3680(&B334&AWT&
:/*-/0*,U&@I!&54&
ContacPng us by email
• Put course number in Subject line: • 6.819 (or 6.869) • Put topic in subject line: Ex: 6.819 Missing Tuesday lecture 6.869 MeePng request: Tuesday 2 pm? 6.819 Late for PS2: Interview
6.869 Project OpPon 2: SuggesPon 6.819 Project OpPon 3: Literature survey
O.</*6>/&,7&,&Z>,-./&<=>?@6<>%A/1#*B%C>6%;2%
Assignments
• Problem sets (60%)
• Final project (40%) – Summary project (5%) – Final presentaPon (5%) – Research component of final project (30%)
• No exams or quizzes
Materials
• Piazza: for student collaboraPon and finding project groups. Instructors and TAs will not be acPve on Piazza.
• Stellar: for turning in late assignments and receiving grades
• Readings: see class website • h=p://6.869.csail.mit.edu/fa15/
Problem sets (60%)
• Four problem sets: 15 % each • CollaboraPon policy – Psets are due individually – Done individually but you can talk to people – WriPng always individually
• Turn a printed version in class. Late due on Stellar. • Up to 4 days late total, for the 4 Psets altogether (i.e. if you use all the 4 days on PS1 for instance, you have none lej for the other PSets).
If you are late ajer that, the grade of the late PS will be zero.
Projects (40%)
Three Project OpPons 1) Summary of final project proposal (5%): 1 page (template) – Individually – Due the first week of November (earlier, be=er!)
2) Research component of final project (30%, template) and final presentaPon (5%). – PresentaPon (2-‐5 minutes each): Dec 3, 8, 10 – Everybody presents.
You are welcome to come to our office hours to brainstorm and suggest your project ideas.
Summary of Project Proposal
• The project proposal should be one page maximum following this template:
• What is the problem/quesKon that you will be invesPgaPng? • What are the most relevant readings? (2-‐4 papers) • What data will you use? • What method or algorithm will you use?
• How will you evaluate your results? QualitaPvely, what kind of results do you expect (e.g. plots or
figures) QuanPtaPvely, what kind of analysis will you use to evaluate and/or
compare your results (e.g. what performance metrics or staPsPcal tests)?
Project : A survey (individual)
• Select a topic (to be discussed with one of us) • Select 10-‐12 papers: send the list • Read the papers • Write a 2500 words survey arPcle (a survey template will be given).
• You can opt for that opPon, change from a coding project to the survey, at any moment before Thanksgiving
Project: Your own project 2-‐4 people
• ApplicaKons/Models. If you have access to a specific large image dataset (e.g. biology, engineering, physics, neuroscience) and a categorizaPon task, you can apply models to this problem.
• From what you learn in class, you can choose a topic/quesPon and propose an approach/model (including quesPons related to neuroscience).
N83b/.7(&E1-1&N>,./0&2<,>>/-Z/&SIH&5/35>/&F&2<,>>/-Z/&!"#$%#&''"(')%*#
Vision: High-Powered Engine
0.5 sec
Brain dynamics of seeing
Human Visual System as a Model
234567,P3-(&>/,8-1-Z&?834&41>>13-0&3?&1-07,-./0%
Aa3&/]78/4/0&3?&+106,>&>/,8-1-Z&
f64;/8&3?&78,1-1-Z&&0,45>/0&
$& $D& $DS& $DC& $DH& $D@&
PQ-.*;/'*K/1%;./3'$2&m/-/8,>1g,P3-&
T1,Z-30P.&?/,768/0&
R1-$.;/'*K/1%;./3'$2%2388/053-*/-./&
d1-*1-Z&7</&*1n/8/-./0&
o&$D!&
#D"DDD"DDD&14,Z/0&p@"DDD&-3-I,;078,.7&-36-0&?834&:38*f/7& M->1-/&14,Z/&0/,8.<&/-Z1-/0&
#D&41>>13-&14,Z/0&
)-*&,j/8&$&U/,8&*3a->3,*1-Z&14,Z/0&&
)"&A388,>;,i&B"&d/8Z60i&:"A"&d8//4,-"&N)EK&SDD#&
)"&A388,>;,i&B"&d/8Z60i&:"A"&d8//4,-"&N)EK&SDD#&
A,8,-P-3&
q17.<.3.[&
E,8U&^6//-&3?&0.370&
`U014,.<1,&-6446>,81,&
T6r/&.3,7&
O368*36Z<&;8/,*&
V30/417/&
N1.3&*/&381g,;,&
`370&&
M?&&
K4,Z/0&
A</&145387,-./&3?&&<,+1-Z&+/-0%&S%R2*L$0%
:<,7&.,-&a/&*3&a17<&,&Z33*&0141>,817U&4/781.&,-*&*%'/-%/S%#*-*_&
Input image Nearest neighbors
The space of world images: As large databases become available, this opens the door to effective data driven methods.
Hays, Efros, Siggraph 2006 Russell, Liu, Torralba, Fergus, Freeman. NIPS 2007`
•! Labels
•! Depth •! …
•! Labels
•! Depth •! …
•! Motion
•! Motion
A</&/+3>6P3-&3?&+1013-&*,7,;,0/0&
# images 103 &&&&&&&105 &&&&&&&109 &&&&&&&107 104 106 108
2MK`ISD&2,>7/.<&$D$&
N)O2)`&GSDD@J&2,>7/.<IH&GSDDCJ&
4
EfKOA&G$%%#J&
S&U/,8&&3>*&[1*&
GSDD%J&
&35$,->,$1-$.$#%#*-*0$-0%
O>1*/&?834&d/1Id/1&`1&L&)-*8/b&X,85,7<U&
O>1*/&?834&d/1Id/1&`1&L&)-*8/b&X,85,7<U&
A35&@&.,7/Z381/0&
A</&/+3>6P3-&3?&+1013-&*,7,;,0/0&
# images 103 &&&&&&&105 &&&&&&&109 &&&&&&&107 104 106 108
# images 103 &&&&&&&105 &&&&&&&109 &&&&&&&107 104 106 108
2MK`ISD&2,>7/.<&$D$&
N)O2)`&GSDD@J&2,>7/.<IH&GSDDCJ&
4
EfKOA&G$%%#J&
S&U/,8&&3>*&[1*&
SUN database (2010) 8 scenes database (2001)
15 scenes database (2006)
base
GSDD%J&
&35$,->,$1-$.$#%#*-*0$-0%
4,$1$>,$1-$.$#%#*-*0$-0%
&&SD$@&
Convolutional Neural Network
23-+3>6P3-& E,]&533>1-Z& f384,>1g,P3-& d6>>U&.3--/.7/*&
O4,87&5<3-/(&5>,./0".0,1>"417"/*6&