ads lab ncku1 michael maire, pablo arbelaez, charless fowlkes, and jitendra malik university of...

17
ADS lab NCKU 1 Michael Maire , Pablo Arbelaez, Charless Fowlkes, and Jitendra Malik university of California, Berkeley – Berkeley university of California, Irvine - Irvine Using Contours to Detect and Localize Junctions in Natural Images 指指指指 : 指指 指指 Reporter: P76961138 指指指 P76964063 指指指 P76964495 指指指

Upload: lora-mathews

Post on 17-Dec-2015

221 views

Category:

Documents


3 download

TRANSCRIPT

Page 1: ADS lab NCKU1 Michael Maire, Pablo Arbelaez, Charless Fowlkes, and Jitendra Malik university of California, Berkeley – Berkeley university of California,

ADS lab NCKU 1

Michael Maire , Pablo Arbelaez, Charless Fowlkes, and Jitendra Malik

university of California, Berkeley – Berkeleyuniversity of California, Irvine - Irvine

Using Contours to Detect and Localize Junctions in Natural Images

指導教授 : 李強 教授Reporter: P76961138 王炳竣

P76964063 方俞淵 P76964495 陳柏嘉

Page 2: ADS lab NCKU1 Michael Maire, Pablo Arbelaez, Charless Fowlkes, and Jitendra Malik university of California, Berkeley – Berkeley university of California,

ADS lab NCKU 2

Outline Introduction flowchart Contour Detection

◦ Local contours◦ Global Contours

Detecting Junctions

Page 3: ADS lab NCKU1 Michael Maire, Pablo Arbelaez, Charless Fowlkes, and Jitendra Malik university of California, Berkeley – Berkeley university of California,

ADS lab NCKU 3

Introduction Contour Detection

Application◦ Image segment

Page 4: ADS lab NCKU1 Michael Maire, Pablo Arbelaez, Charless Fowlkes, and Jitendra Malik university of California, Berkeley – Berkeley university of California,

ADS lab NCKU 4

Introduction Junctions Detection

Application

Page 5: ADS lab NCKU1 Michael Maire, Pablo Arbelaez, Charless Fowlkes, and Jitendra Malik university of California, Berkeley – Berkeley university of California,

ADS lab NCKU 5

flowchart

Image

Local contours

Globalcontours

Result

Contour Detection:

Junctions Detection:

Contour DetectionResult

Result

Page 6: ADS lab NCKU1 Michael Maire, Pablo Arbelaez, Charless Fowlkes, and Jitendra Malik university of California, Berkeley – Berkeley university of California,

ADS lab NCKU 6

Contour Detection: Local contours

Boundary Cues

Model

Pb

Brightness

Color

Texture

Goal: learn the posterior probability of a boundary Pb(x,y,) from local information only

Cue Combination

Martin, Fowlkes, Malik PAMI 04

Page 7: ADS lab NCKU1 Michael Maire, Pablo Arbelaez, Charless Fowlkes, and Jitendra Malik university of California, Berkeley – Berkeley university of California,

ADS lab NCKU 7

Contour Detection: Local contours

Line Detection

梯度運算子 (Gradient Operator)

Page 8: ADS lab NCKU1 Michael Maire, Pablo Arbelaez, Charless Fowlkes, and Jitendra Malik university of California, Berkeley – Berkeley university of California,

Individual Features Brightness Gradient BG (x,y,r,)

Color Gradient CG (x,y,r,)

Texture Gradient TG (x,y,r,)

ADS lab NCKU 8

Contour Detection: Local contours

These are combined using logistic regression

r

(x,y)

Page 9: ADS lab NCKU1 Michael Maire, Pablo Arbelaez, Charless Fowlkes, and Jitendra Malik university of California, Berkeley – Berkeley university of California,

ADS lab NCKU 9

Contour Detection: Global Contours

Node : the points in the feature space

Graph: G(V , E)

Edge : between every pair of points

Weight(Wij)

Image segmentation (high-level cues)

Page 10: ADS lab NCKU1 Michael Maire, Pablo Arbelaez, Charless Fowlkes, and Jitendra Malik university of California, Berkeley – Berkeley university of California,

How to partition the Graph:

ADS lab NCKU 10

Contour Detection: Global Contours

Partition graph so that similarity within group is large and similarity between groups is small -- Normalized Cuts [Shi & Malik 97]

Page 11: ADS lab NCKU1 Michael Maire, Pablo Arbelaez, Charless Fowlkes, and Jitendra Malik university of California, Berkeley – Berkeley university of California,

ADS lab NCKU 11

Contour Detection: Global Contours

Page 12: ADS lab NCKU1 Michael Maire, Pablo Arbelaez, Charless Fowlkes, and Jitendra Malik university of California, Berkeley – Berkeley university of California,

ADS lab NCKU 12

Contour Detection: Global Contours

Page 13: ADS lab NCKU1 Michael Maire, Pablo Arbelaez, Charless Fowlkes, and Jitendra Malik university of California, Berkeley – Berkeley university of California,

ADS lab NCKU 13

Contour Detection: Combine

Global ContoursLocal contours

Page 14: ADS lab NCKU1 Michael Maire, Pablo Arbelaez, Charless Fowlkes, and Jitendra Malik university of California, Berkeley – Berkeley university of California,

1.Estimate the optimal junction location

by minimizing its weighted distance from the contours

2.Update the weight of each contour

in order to select only those contours passing close to the junction

ADS lab NCKU 14

Detecting Junctions

Page 15: ADS lab NCKU1 Michael Maire, Pablo Arbelaez, Charless Fowlkes, and Jitendra Malik university of California, Berkeley – Berkeley university of California,

ADS lab NCKU 15

Detecting Junctions

Page 16: ADS lab NCKU1 Michael Maire, Pablo Arbelaez, Charless Fowlkes, and Jitendra Malik university of California, Berkeley – Berkeley university of California,

ADS lab NCKU 16

Demo

Page 17: ADS lab NCKU1 Michael Maire, Pablo Arbelaez, Charless Fowlkes, and Jitendra Malik university of California, Berkeley – Berkeley university of California,

ADS lab NCKU 17

Conclusions & Comment 上課時呈現 ^__^