Linear SVM Classification of Visual Objects using a Dense Image Representation

1
Linear SVM Classification of Visual Objects
using a Dense Image Representation

• Diane Larlus, Gyuri Dorko, Bill Triggs and
  Frederic Jurie INRIA Rhone-Alpes, 665, avenue de
  l'Europe
• 38330 Montbonnot, FRANCE
• Email frederic.jurie_at_inrialpes.fr

2
Overview

• Similarities with the previous approach
• Geometry free method
• Based on codebook, used to encode images
• Differences
• Dense representation of images for building
  construction interest point operators often
  failed on small objects.
• Images features (two different representations
  are used)
• Linear classifier

3
Codebook construction
Training Images
Quantization

• Feature space
• Raw pixel intensities (normalized) 15x15 windows
• SIFT description (128-d)

4
Image representation

• Images Representations
• (one vector per image)
• histograms
• Binary vectors

Training Images
Quantization
5
Dimensionality reduction

• Odd ratio
• Mutual information
• Gain in mutual information
• Svm based dimensionality reduction

6
Image classification

• Classifiers
• Naive Bayes
• SVM
• Linear
• Polynomial
• RBF

-
Binary classifiers
7
Keypoint Clustering the feature space

• Large number of vectors 10 millions
• K-mean gives bad results, because of unbalance in
  cluster sizes
• low of power, known as Zipfs law in text-based
  word frequency analysis

8
Facing unbalanced densities

• Well know problem, solutions exist
• Sub-sampling
• Over-sampling
• Re-weighting
• We developed an on-line method based on
  sub-sampling
• uniform sampling of vectors
• produce largest clusters
• Re-sample the set of vectors (without sampling
  in existing clusters)
• On-line stop criterion (nb. of clusters, quality
  of the quantization, quality of the
  classification task , etc.)

9
Experimentswith the PASCAL dataset (test1 only)

• Classifiers
• Naive Bayes
• SVM
• Linear
• Polynomial
• RBF
• -gtparameters

• Feature space
• Raw pixel intensities (normalized) 15x15 windows
• SIFT description (128-d)
• -gtparameters

Image representation - binary vectors -
histograms

• Codebook
• Size
• Dimensionality reduction?
• Parameters of the clusterer

10
(No Transcript)
11
(No Transcript)
12
(No Transcript)
13
(No Transcript)
14
cars
15
bikes
16
motorbikes
17
people
18
Comments on the results

• Binary representation vs. histogram we are not
  capturing texture but occurrences of
  informative parts.
• Linear SVM
• Dimensionality
• Dense vs. sparse

19
Conclusions

• In case of dense representation
• Feature representation
• Dimensionality reduction
• Image representation
• Classifier
• Are not critical issues, the most importantissue
  is the codebook and the way it is computed.
• Linear SVM / non-linear SVM
• Perspective
• is it possible to go further with this type of
  approach ?
• Is it the right way to evaluate
  theclassification task (role of thebackground)
  we should have background only images.

This is a car image!
20
Suggestion for next competitions

• Choose more carefully training / test images
• Annotate training images more accurately (20 of
  120 motorbike images are labeled as motorbike
  only while then contains humans too)
• Test image should be related to training images
• Only allow to train object vs. background,
  provide background images
• Evaluation test images should have weights
  (related to the visibility of objects).
  Capability of detecting visible objects (of not
  missing them) capability of detecting difficult
  objects.