Object class recognition using unsupervised scaleinvariant learning

1
EC Project Pascal Object recognition challenge
Andrew Zisserman Luc Van Gool Chris Williams
2
Two objectives

• Service
• collate datasets and annotation/ground truth
  access functions
• Standardization
• define performance measurement tests

• 4 person-months
• October-December 2004

3
Category data base

• Explore dimensions of variation
• scale/position within image
• lighting
• background clutter
• pose/visual aspect
• partial occlusion
• within class variation
• deformation
• texture
• articulation
• inter class variation
• This requires many image instances

4
For each database

• include a description of level of difficulty in
  terms of
• size variation pose variation background
  clutter occlusion intra-class variation etc
• source of database (lab and if any restrictions
  e.g. from Google)
• acknowledgments
• relevant publications

• Matlab access functions
• read image into matlab
• ground truth region of interest read from
  text/XML data file
• rectangle for each instance in image

5
e.g. Caltech Motorbikes
6
Other datasets The Caltech six background
7
Graz datasets Opelt/Pinz

• Several hundred images at varying levels of
  difficulty for each of
• people
• cars
• bikes (pedal cycles)
• neutral backgrounds

8
101 Categories (Caltech)

• Collected from Google
• 40-800 images per category
• Chosen at random from picture dictionary

9
(No Transcript)
10
Cow Database (Leibe Schiele)

• Original Data
• 60 cow sequences (from Derek Magee _at_ Leeds)
• Training Data
• Extracted frames from subset of sequences
• Train on 113 images (segmented)
• Tests on novel sequences
• 14 sequences, 2217 frames total
• 18 unseen cows, different backgrounds
• 1682 instances of cows gt50 visible

11
Motorbikes - from the Web
12
Standardization

• Particular selection of images for training /
  testing, e.g.
• those used by Agarwal and Roth ECCV 2002
• those used by Fergus et al in CVPR 2003
• ones that explore only intra-class variation
  (but not scale, orientation, visual aspect etc)