Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

21
Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460

Transcript of Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

Page 1: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

Pulkit AgrawalY7322

BVV Sri Raj DuttY7110

Sushobhan NayakY7460

Page 2: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

OutlineWhat is a sceneScene recognitionMethodResultsFuture WorkReferences

Page 3: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

What is a Scene?Scene- as opposed to

‘object’ or ‘texture’

Object: when view subtends 1 to 2 meters around observer---hand distance

Page 4: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

What is a Scene?

observer and fixated point- >5 meters

Page 5: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

Scene Recognition2 approachesObject recognition

Global info – details and object info ignoredo Experimental

evidenceo ‘Gist’ of image

Page 6: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

Scene RecognitionExclusive

classificationStructural

attributes- Continuous organization of scenes along semantic axes

Page 7: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

Semantic axes2 levels:

Degree of naturalness: man-made to natural landscape

Ambiguous (building in field) pictures around center

Page 8: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

Semantic axesNatural scenes-

degree of openness

Artificial urban scenes- degree of verticalness and horizontalness

Highways--Highways +Tall Building---Tall Buildings

Page 9: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

Method

Information at various Scales

What do we Need ??

High Frequency ? Low Frequency ?

Both ??

Page 10: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

Feature ExtractionImage Power Spectrum

Gabor Filters (Scale, Orientation)

Features (512 used)

Page 11: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

Mathematical Details…Important data from Image power spectrum

Structural discriminant feature

DST=Discriminat Spectral Template- --an encoding of the discriminant structure between two image categories

‘u’ -weighted integral of power spectrum

Page 12: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

Classification

Image = Feature Vector()

Required Classes

Linear Discriminant Analysis

Discriminating Vector (D.V)Maximum Separation b/w classes

Page 13: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

Mathematical Details…..Image represented as Feature Vector x.m1 , m2: mean vector of feature vector of 2

classes

Page 14: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

Mathematical Details…

gn= feature

Gn = Gabor filter

dn = through learning

Page 15: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

Learning…Projection of Training Set

Image F.V. on D.V.

Use LDA to determine Threshold

Classifier Obtained

Page 16: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

Learning

Page 17: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

Work..

Artificial v/s Natural

Open v/s Non Open

Page 18: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

ResultsArtificial v/s Natural

Artificial•80 Test Images•67 classified Correctly

Natural•80 Test Images•75 classified Correctly

89% Correct results

Page 19: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

Result

Page 20: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

Future WorkArrangement in semantic axesAddition of features

Depth Symmetry

Contrast Ruggedness

8 category arrangement (skyscrapers, highway, street, flat building, beach, field, mountain, forest)

Experiment with Haar and other filters

Page 21: Pulkit Agrawal Y7322 BVV Sri Raj Dutt Y7110 Sushobhan Nayak Y7460.

ReferenceTorralba A. & Olivia A., Semantic

Organisation of Scenes using Discriminant Structural Templates (1999)

Torralba A. & Olivia A., Modeling the Shape of the Scene: A Holistic Representation of the Spatial Envelope(2001)

Olivia A., Gist of the Scenehttp://people.csail.mit.edu/torralba/code/spati

alenvelope/