Alvaro Soto

Transcripción

Alvaro Soto
DCC - PUC
Case Study
80 million tiny images: a large dataset for nonparametric object and scene recognition
A. Torralba, R. Fergus and W. T. Freeman
PAMI, Nov. 2008
Alvaro Soto
Alvaro Soto
1
DCC - PUC
K-Vecinos Cercanos
Predicción
Uso de métrica
de distancia
Clasificación
Votación por
mayoría
Alvaro Soto
2
DCC - PUC
Parámetros y Modelos
• Modelos Paramétricos
• Ej. Regresión lineal
• Modelos SemiParamétricos
• Ej. Fourier, wavelets
• Modelos No Paramétricos
• Ej. K-vecinos cercanos
Alvaro Soto
3
DCC - PUC
What can you do with 80M images?
A. Torralba, R. Fergus and W. T. Freeman. 80 million tiny images: a large dataset for non-parametric object and
Alvaro Soto
scene recognition. IEEE PAMI, 2006.
4
DCC - PUC
Alvaro Soto
5
DCC - PUC
Alvaro Soto
6
DCC - PUC
Alvaro Soto
7
DCC - PUC
Alvaro Soto
8
DCC - PUC
How to measure similarity between images?
Patch horizontal mirror,
translations and scaling.
Individual pixel shift
in 5x5 window. T given
By best Dwarp.
Alvaro Soto
9
DCC - PUC
Image matching using distance metrics:
Dssd , Dwarp and Dshift
Alvaro Soto
10
DCC - PUC
Images
They use 7 independent image search engines: Altavista, Ask, Flickr, Cydral,
Google, Picsearch and Webshots.
They extract all non-abstract nouns from wordnet: 75.846 in total.
They automatically download all the images provided by each engine for all
non-abstract nouns.
Running over 8 months, this method gathered 97.245.098 images in total.
Once intra-word duplicates and uniform images (images with zero variance) are
removed, this number is reduced to 79.302.017 images from 75.062 words
(around 1% of the keywords had no images).
They store the images using a resolution of 32x32 pixels (eficiency).
Alvaro Soto
11
DCC - PUC
Wordnet
Wikipedia: WordNet es una enorme base de datos léxica del idioma inglés. Agrupa
las palabras en conjuntos de sinónimos llamados 'synsets', almacenando las relaciones
semánticas entre estos conjuntos de sinónimos.
Alvaro Soto
12
DCC - PUC
A histogram of images per keyword collected
Alvaro Soto
13
DCC - PUC
Precision in Outputs from Search Engines
Accuracy drops after the 100th image and then stabilizes at
around 44% correct on average.
Alvaro Soto
14
DCC - PUC
Accuracy of labeling
for different nodes
of a portion of the
Wordnet tree.
More precision
for more specific
words
Alvaro Soto
15
DCC - PUC
Recognition Using Wordnet and Original
Keywords Used to Retrieve Each Image
Alvaro Soto
16
DCC - PUC
Alvaro Soto
17

Documentos relacionados