56
Filters and other potions P. Perona - Caltech MIT - 21 November 2013

Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

  • Upload
    vanbao

  • View
    217

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Filters and other potionsP. Perona - Caltech

MIT - 21 November 2013

Page 2: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture
Page 3: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

? whatwhere

Page 4: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Architectures

Page 5: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Architecture 1

Image(s)

The vision black box

Ripe bananas

Marble torso

train

build

ing

Feature extraction: texture

stereo disparity color contrast motion flow

edgels ….

Surface shape, scene depth,

spatial relationships, 3D motion

Grouping: image regions

Perceptual organization: 2.5D sketch:

boundaries, junctions, foregrnd, bckgrnd

Recognition, surface properties

Image processing Regions and surfacesObjects, verbs, categories…

motor

cognition

[Marr ’82]

Page 6: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

features?

Le Corbusier, Villa Savoyehttp://flickr.com/photos/ikura/1398271367/

Page 7: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

edges

http://www.iit.edu/~stawraf/perspx.jpgLe Corbusier, Villa Savoye

Page 8: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture
Page 9: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture
Page 10: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture
Page 11: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

[Fukushima ‘80]

Architecture 2

Page 12: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture
Page 13: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

[DeValois ’85]

Page 14: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Column

Page 15: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Hypercolumn

Page 16: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture
Page 17: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture
Page 18: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Dense sampling

Page 19: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

translation, rotation invariance

[LeCun et al. 1998]

Page 20: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

scale invariance

[Lowe 2004]

Page 21: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

[Hinton et al. ’12]

translation, rotation, scale invariance

Page 22: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

96 filters 6 orientations 2 center-surround 14 scale samples over 2.2 binary octaves

Page 23: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Detection Performance

Caltech pedestrians: 1M frames, 250K hand-annotated

Page 24: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Detection Performance

Page 25: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Detection Performance

Dollar et al. ‘10Dollar et al. ‘08

Viola & Jones ‘01

Dalal-Triggs ‘05 *

Walk et al. ‘10

Page 26: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

filter technology

Page 27: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Scale, orientation, elongation…. lots of CPU cycles

Page 28: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

how do we make computations efficient?

Page 29: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Separability

[Adelson & Bergen, ’85]Cost = m x n Cost = m + n

R(i, j) =X

h=1:M,k=1:N

k(h, k)I(i� h, j � k) R(i, j) =X

h=1:M

X

k=1:N

k(h)k0(k)I(i� h, j � k)

Page 30: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Separability and decomposition

[Adelson & Bergen, ’85]

Page 31: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Steerability

[Freeman & Adelson, ’91]

Page 32: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

General decomposition

k(x, ✓) =DX

i=1

bi(✓)fi(x)

k(x, y) =DX

i=1

fi(x)gi(y)

k(x, y; ✓) =DX

i=1

bi(✓)fi(x)gi(y)

Page 33: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Design?

Page 34: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

x

=k(x; ✓)

D

bi(✓)

x

�i,i

fi(x)

A = USV T

Page 35: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

ApproximationK(x, y; ✓) =

DX

i=1

bi(✓)fi(x, y)

K(x, y; ✓) ⇡RX

i=1

bi(✓)fi(x, y) R ⌧ D

Page 36: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture
Page 37: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture
Page 38: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

[Perona ’95]

Page 39: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

[Perona ’95]

Page 40: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

[Perona ’95]

Page 41: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Tensor Factorization

k(x, y; ✓) =DX

i=1

bi(✓)fi(x)gi(y)

•Not a convex problem •Gradient descent

[Shy, Perona ’96]

Page 42: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Including scale by resampling

Page 43: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture
Page 44: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

[Manduchi et al. ’98][cfr. Simoncelli et al]

Page 45: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture
Page 46: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Exploiting Image Statistics

Page 47: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

original

upsampled

sampling the gradient

Page 48: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

[Dollar et al. 2013]

Page 49: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Gradient histograms[Dollar et al. 2013]

Page 50: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Power law feature scaling

Page 51: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Power law feature scaling

Page 52: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Individual images

[Dollar et al. 2013]

Page 53: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Fast computations

Page 54: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Fast computations

[Dollar et al. 2013]

Page 55: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Performance

[Dollar et al. 2013]

Page 56: Filters and other potions - LCSLlcsl.mit.edu/ldr-workshop/Slides/Perona_LDR_MIT_112213.pdf · potions P. Perona - Caltech MIT - 21 November 2013? what where. Architectures. Architecture

Conclusions• Filtering front-end

• Need fine sampling of scale, orientation, …

• Scalable, separable and steerable approximations

• Exploiting image statistics to extrapolate

• Fast and accurate detection