Challenges to image parsing
        researchers
         Lana Lazebnik
        UNC Chapel Hill
                                             sky



                                             mountain
                           building




                  person
                                car          car
                  sidewalk
                                      road
The past: “closed universe”
   datasets
   Tens of classes, hundreds of images, offline learning




                                                               Figure from Shotton et al. (2009)

He et al. (2004), Hoiem et al. (2005), Shotton et al. (2006, 2008, 2009), Verbeek and Triggs
(2007), Rabinovich et al. (2007), Galleguillos et al. (2008), Gould et al. (2009), etc.
The future: “open universe”
datasets
       Evolving images, annotations




            http://labelme.csail.mit.edu/
The future: “open universe”
                          datasets
                                 Non-uniform class frequencies
                     12
Millions of Pixels




                     10
                      8
                      6
                      4
                      2
                      0
Which “closed universe”
techniques can survive in the
“open universe” setting?
   Combination of local cues?
   Multiple segmentations/grouping hypotheses?
   Context?
   Graphical models (MRFs, CRFs, etc.)?
   Offline learning and inference?
Learning from all of LabelMe
50K images, 232 labels

                                                    sky
             window                    building

                          building     car                     car
             door                              road
                                                  sidewalk
                road

                    sky
                                         sky


                          tree car
                                                    building

                road                  mountain


                           sun               ceiling
              sky

                                                       wall

                     sea
                                               floor


                                     Tighe & Lazebnik, work in progres
Learning from all of LabelMe
       50K images, 232 labels

                     Per-class classification rates
                      SiftFlow   Barcelona   LM + Sun
100%
 75%
 50%
 25%
  0%




100%
 75%
 50%
 25%
  0%




                                                        Tighe & Lazebnik, work in progres
Challenge: Parsing high-res
images
Challenge: Dynamic image
interpretation
   Image parsing algorithms should become
    autonomous decision-making agents


                                Visual “detective
                                task”: Where
                                was this photo
                                taken?
Challenge: Dynamic image
interpretation
   Image parsing algorithms should become
    autonomous decision-making agents
Input
Summary
   Challenges to image parsing researchers:
     Learn  to parse images from “open universe”
      evolving datasets
     Try parsing gigapixel images!

     Develop active, sequential image interpretation
      strategies

Fcv scene lazebnik

  • 1.
    Challenges to imageparsing researchers Lana Lazebnik UNC Chapel Hill sky mountain building person car car sidewalk road
  • 2.
    The past: “closeduniverse” datasets Tens of classes, hundreds of images, offline learning Figure from Shotton et al. (2009) He et al. (2004), Hoiem et al. (2005), Shotton et al. (2006, 2008, 2009), Verbeek and Triggs (2007), Rabinovich et al. (2007), Galleguillos et al. (2008), Gould et al. (2009), etc.
  • 3.
    The future: “openuniverse” datasets Evolving images, annotations http://labelme.csail.mit.edu/
  • 4.
    The future: “openuniverse” datasets Non-uniform class frequencies 12 Millions of Pixels 10 8 6 4 2 0
  • 5.
    Which “closed universe” techniquescan survive in the “open universe” setting?  Combination of local cues?  Multiple segmentations/grouping hypotheses?  Context?  Graphical models (MRFs, CRFs, etc.)?  Offline learning and inference?
  • 6.
    Learning from allof LabelMe 50K images, 232 labels sky window building building car car door road sidewalk road sky sky tree car building road mountain sun ceiling sky wall sea floor Tighe & Lazebnik, work in progres
  • 7.
    Learning from allof LabelMe 50K images, 232 labels Per-class classification rates SiftFlow Barcelona LM + Sun 100% 75% 50% 25% 0% 100% 75% 50% 25% 0% Tighe & Lazebnik, work in progres
  • 8.
  • 9.
    Challenge: Dynamic image interpretation  Image parsing algorithms should become autonomous decision-making agents Visual “detective task”: Where was this photo taken?
  • 10.
    Challenge: Dynamic image interpretation  Image parsing algorithms should become autonomous decision-making agents Input
  • 11.
    Summary  Challenges to image parsing researchers:  Learn to parse images from “open universe” evolving datasets  Try parsing gigapixel images!  Develop active, sequential image interpretation strategies

Editor's Notes

  • #4 Challenges for large-scale learning: