Scribd
Upload
55 views36 pages

Week11 FinalDeck PDF

The document discusses different approaches for object recognition and detection in images, including: - Image classification which identifies class labels and probabilities for an entire image - Object detection which finds specific objects and their locations within an image - Early approaches used global representations and holistic matching of features - Current methods often use local features with machine learning models like deformable part models (DPM) which represent objects as a root template plus flexible parts

Uploaded by

Sagor Saha
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
55 views36 pages

Week11 FinalDeck PDF

The document discusses different approaches for object recognition and detection in images, including: - Image classification which identifies class labels and probabilities for an entire image - Object detection which finds specific objects and their locations within an image - Early approaches used global representations and holistic matching of features - Current methods often use local features with machine learning models like deformable part models (DPM) which represent objects as a root template plus flexible parts

Uploaded by

Sagor Saha
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 36

Video 11.

1
Kostas Daniilidis

Property of University of Pennsylvania, Kostas Daniilidis 1


By object recognition we might
ea …
Image classification:
Input is an image
and output is a set of
class labels with
probabilities

https://cloud.google.com/vision/docs/drag-and-drop

Property of University of Pennsylvania, Kostas Daniilidis 2


.. Or here:

Property of University of Pennsylvania, Kostas Daniilidis 3


But object recognition usually
ea s…

Object Detection
and Localization:
Find where is a car
in the image?

Property of University of Pennsylvania, Kostas Daniilidis 4


Property of University of Pennsylvania, Kostas Daniilidis 5
History (Grauman and Leibe, 2011)

Property of University of Pennsylvania, Kostas Daniilidis 6


Recognition ingredients

• Object features (for example, HOG) that is resistent


to geometric and photometric nuisances.
• Learning (training) the class model (SVM) to
represent intra-class variations.
• Testing on an input image and output all bounding
boxes for a specific class (Sliding window)

Property of University of Pennsylvania, Kostas Daniilidis 7


Instance recognition by matching (Lowe, 2004)

Match SIFT Features and verify geometric consistency

Property of University of Pennsylvania, Kostas Daniilidis 8


Video 11.2
Kostas Daniilidis

Property of University of Pennsylvania, Kostas Daniilidis 9


Holistic approaches (Grauman and Leibe, 2011)

Property of University of Pennsylvania, Kostas Daniilidis 10


Class Recognition: Sliding
windows
• Learn a HOG representative of • What happens if we slide this
a car over multiple scales window over an input image?

Property of University of Pennsylvania, Kostas Daniilidis 11


HOG filter

Property of University of Pennsylvania, Kostas Daniilidis 12


Property of University of Pennsylvania, Kostas Daniilidis 13
Holistic / Global Representation
+ Gestaltism: shape is
perceived as a whole:
- not merely sum of parts
- global dependencies

Where is the bottle?


clutter > 85%
of edges

- Susceptibility to clutter
?
=
edges model
1
12/14/10 Alexander Toshev
4

Property of University of Pennsylvania, Kostas Daniilidis 14


Localization through Shape Verification

Superpixel segmentation of DPM hypotheses and


selection of superpixels that best match our shape
representation

15

Property of University of Pennsylvania, Kostas Daniilidis 15


Holistic approaches need perceptual grouping

Property of University of Pennsylvania, Kostas Daniilidis 16


Property of University of Pennsylvania, Kostas Daniilidis 17
Property of University of Pennsylvania, Kostas Daniilidis 18
Property of University of Pennsylvania, Kostas Daniilidis 19
Property of University of Pennsylvania, Kostas Daniilidis 20
Property of University of Pennsylvania, Kostas Daniilidis 21
Chordiogram vs Shape Context
chordiogram shape context

pairs w.r.t.
all pairs
to a fixed offset

Translation invariance:Yes No

Global support: Yes No

Global configuration: No Yes

Notion of parts: No Yes

Property of University of Pennsylvania, Kostas Daniilidis 22


Figure/Ground Organization
Chord normals point
towards the object
interior.

Contour and interior


descriptor:
• More discriminative than pure
contour descriptors.
• Encode figure/ground
organization.
captures face captures vase
Rubin’s vase

12/14/10 Alexander Toshev

Property of University of Pennsylvania, Kostas Daniilidis 23


24

Property of University of Pennsylvania, Kostas Daniilidis 24


Video 11.3
Kostas Daniilidis

Property of University of Pennsylvania, Kostas Daniilidis 25


Deformable Part Models

• DPM consists a Root and several


Parts
• Root represents holistic object shape
• Part captures detailed part appearance

A car DPM m od el

Root Parts

Property of University of Pennsylvania, Kostas Daniilidis 26


Deformable Part Models

• DPM consists a Root and several


Parts
• Root represents holistic object shape
• Part captures detailed part
appearance
A car DPM m od el

Parts
Root

Property of University of Pennsylvania, Kostas Daniilidis 27


Deformable Part Models

• Flexible part configuration


• Parts are allowed to move around the
anchor points (default position)

Deform ation Parts

Property of University of Pennsylvania, Kostas Daniilidis 28


Property of University of Pennsylvania, Kostas Daniilidis 29
Deformable Part Models

• Detector response at a given root


location

Root Parts bias

Property of University of Pennsylvania, Kostas Daniilidis 30


Property of University of Pennsylvania, Kostas Daniilidis 31
Property of University of Pennsylvania, Kostas Daniilidis 32
Property of University of Pennsylvania, Kostas Daniilidis 33
Property of University of Pennsylvania, Kostas Daniilidis 34
Property of University of Pennsylvania, Kostas Daniilidis 35
Property of University of Pennsylvania, Kostas Daniilidis 36

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy