1. Modeling the Shape of the
Scene: A Holistic
Representation of the Spatial
Envelope
Oliva. A., Torralba A. (2001).
International Journal of Computer
Vision 42(3), 145-175.
2. Scene Recognition
- What is a Scene?
- S.R. Studies
Spatial Envelope
- Holistic
Representation
Scene Recognition
- Definition
- Spatial Properties
Scene Structure 2 different approaches:
- DFT
- WFT Details
General Picture
- Mean
Spectrograms General Picture Details
Sp.E. Properties
- DST
- WDST
- Naturalness Scene Recognition is…
- Natural Scenes’
Properties a reconstruction of the properties of a 3D
- Urban Scenes’
Properties Scene that are important for tasks such as
- Correlations
- Semantic
movement or grasping
Categories the providing of information about the
Conclusions
semantic category and the function of the
environment.
3. Scene Recognition
- What is a Scene?
- S.R. Studies
Spatial Envelope
- Holistic
Representation
Scene Recognition
- Definition
- Spatial Properties
Scene Structure Object Scene
- DFT
- WFT 1 – 2 meters >5meters
- Mean
Spectrograms
“Hand distance” “Movement”
Sp.E. Properties
- DST
- WDST
- Naturalness
- Natural Scenes’
Properties
- Urban Scenes’
Properties
- Correlations
- Semantic
Categories
Conclusions
4. Scene Recognition
- What is a Scene?
- S.R. Studies
Spatial Envelope
- Holistic
Representation
Scene Recognition
- Definition
- Spatial Properties
Scene Structure “Gist”
Recognition Speed: 200ms
- DFT
- WFT
- Mean View of a Scene for short time
Spectrograms
Sp.E. Properties
- DST
Recognition of Functional and
- WDST
- Naturalness Categorical Properties
- Natural Scenes’
Properties Object information – ignored
- Urban Scenes’
Properties
- Correlations Environmental pictures – rapid
- Semantic
Categories
Conclusions categorization
5. Scene Recognition
- What is a Scene?
- S.R. Studies
Spatial Envelope
- Holistic
Representation
Spatial Envelope
- Definition
- Spatial Properties
Scene Structure Scene is not just an arrangement of objects and
- DFT
- WFT
it’s something more than the relationship
- Mean between them.
Spectrograms
Sp.E. Properties
- DST
- WDST
- Naturalness
- Natural Scenes’
Properties
- Urban Scenes’
Properties
- Correlations
- Semantic
Categories
Conclusions
Scenes of the same category have a
stable spatial structure.
6. Scene Recognition
- What is a Scene?
- S.R. Studies
Spatial Envelope
- Holistic
Representation
Spatial Envelope
- Definition
- Spatial Properties
Scene Structure Spatial Envelope: boundaries that give a shape to the Scene.
- DFT
- WFT
e.g. ground, horizon, elevations, slants
- Mean
Spectrograms
Sp.E. Properties
- DST
- WDST
- Naturalness
- Natural Scenes’
Properties
- Urban Scenes’
Properties
- Correlations
- Semantic
Categories
Conclusions
7. Scene Recognition
- What is a Scene?
- S.R. Studies
Spatial Envelope
- Holistic
Representation
Spatial Envelope
- Definition
- Spatial Properties
Scene Structure
81 pictures
- DFT
- WFT Experiment A:
- Mean
Spectrograms 17 subjects - 81 pictures
Sp.E. Properties Criteria: A B
- DST
1)Global Aspect
- WDST
- Naturalness 2)Global Structure
- Natural Scenes’ 3)Elements A1 A2 B1 B2
Properties
- Urban Scenes’
Properties
a b a b a b a b
- Correlations
- Semantic
Categories 8 subgroups
Conclusions
8. Scene Recognition
- What is a Scene?
- S.R. Studies
Spatial Envelope
- Holistic
Representation
Spatial Envelope
- Definition 5 Main Properties
- Spatial Properties
Scene Structure
1. Naturalness
- DFT sharp VS smooth edges
- WFT
- Mean
Spectrograms 1. Openness
Sp.E. Properties
- DST No Horizon / Visual references
- WDST VS Horizon / lack of Visual
- Naturalness References
- Natural Scenes’
Properties
- Urban Scenes’ 1. Roughness
Properties
- Correlations Complexity VS No complexity
- Semantic
Categories
Conclusions 1. Expansion
Depth VS No depth
1. Ruggedness
Deviation VS No deviation
9. Scene Recognition
- What is a Scene?
- S.R. Studies
Spatial Envelope Modeling the Scene Structure
Image-based representations
- Holistic
Representation
- Definition
Hanning
- Spatial Properties Discrete Fourier Transform Window
Scene Structure Intensity
- DFT
- WFT
- Mean
Spectrograms
Sp.E. Properties
- DST
- WDST Amplitude Phase
- Naturalness Spectrum Function
- Natural Scenes’
Properties
- Urban Scenes’
Properties Windowed Fourier Transform
- Correlations
- Semantic
Categories
Hamming
Conclusions Intensity window with
radius r
10. Scene Recognition
- What is a Scene?
- S.R. Studies
Spatial Envelope
- Holistic
Representation
Modeling the Scene Structure
- Definition Energy Spectrum
- Spatial Properties
Scene Structure
- DFT
- WFT
- Mean
Spectrograms
Sp.E. Properties
- DST
- WDST
- Naturalness Energy Spectrum Principal Components of Energy Spectrum
- Natural Scenes’
Properties
- Urban Scenes’
Spectrogram
Properties
- Correlations
- Semantic
Categories
Conclusions
Spectrogram Principal Components of the Spectrogram
11. Scene Recognition
- What is a Scene?
- S.R. Studies
Spatial Envelope
Modeling the Scene Structure
- Holistic
Representation
- Definition
- Spatial Properties
Scene Structure
- DFT
- WFT
- Mean
Spectrograms
Sp.E. Properties Man-made open urban | Streets City Center Buildings close-up
- DST Urban Str.
- WDST
- Naturalness
- Natural Scenes’
Properties
- Urban Scenes’
Properties
- Correlations
- Semantic
Categories
Natural open natural enclosed mountains enclosed forest natural,
Conclusions
non-textured
12. Scene Recognition
- What is a Scene?
- S.R. Studies
Spatial Envelope
Estimation of the Spatial Envelope
- Holistic
Representation Properties
- Definition
- Spatial Properties Scene
Scene Structure attribute s
- DFT
- WFT
- Mean
Spectrograms
Sp.E. Properties
- DST
- WDST
- Naturalness
- Natural Scenes’
Properties
- Urban Scenes’
Properties
- Correlations
- Semantic
Categories
Conclusions
Discriminant Spectral Template - DST
13. Scene Recognition
- What is a Scene?
- S.R. Studies
Spatial Envelope
Estimation of the Spatial Envelope
- Holistic
Representation Properties
- Definition
- Spatial Properties
Scene Structure
- DFT
- WFT
- Mean
Spectrograms
Sp.E. Properties
- DST
- WDST
- Naturalness
- Natural Scenes’
Properties
- Urban Scenes’
Properties
- Correlations
- Semantic
Categories
Conclusions
Windowed Discriminant Spectral Template -
WDST
14. Scene Recognition
- What is a Scene?
- S.R. Studies
Spatial Envelope
Spatial Envelope Properties
- Holistic
Representation ~Naturalness~
- Definition
- Spatial Properties
Scene Structure 2000 images of natural and man-made scenes
- DFT
- WFT
- Mean
Spectrograms
Sp.E. Properties
- DST
- WDST
- Naturalness
DST(fx,fy) WDST(fx,fy)
- Natural Scenes’
Properties
- Urban Scenes’
93.5% Correct!
Properties
- Correlations
- Semantic
Categories
Conclusions
Low Naturalness High Naturalness
22. Scene Recognition
- What is a Scene?
- S.R. Studies
Spatial Envelope
- Holistic Conclusions / Important Points
Representation
- Definition
- Spatial Properties
Scene Structure
•Introduction of the “Spatial Envelope” model
- DFT •A set of important perceptual properties (Naturalness, openness,
- WFT ruggedness, expansion, roughness)
- Mean
Spectrograms •Provides a holistic description of the scene
Sp.E. Properties
- DST
•Spatial Envelope properties are strongly correlated with
- WDST the templates of DST and WDST
- Naturalness
- Natural Scenes’
•The S.E. model organizes scenes as humans subjects do
Properties
- Urban Scenes’ •S.E. model is able to retrieve images sharing the same
Properties semantic category
- Correlations
- Semantic •Object identification is not a necessary stage for
Categories
identification a scene
Conclusions
•S.E. model provides a meaningful representation of
complex scenes