Semantic Robot Vision Challenge: Current State and Future Directions

1
Semantic Robot Vision Challenge Current State
and Future Directions

• Scott Helmer, David Meger, Pooja Viswanathan,
  Sancho McCann, Matthew Dockrey, Pooyan Fazli,
  Tristram Southey, Marius Muja, Michael Joya, Jim
  Little, David Lowe, Alan Mackworth

2
What is the point of robotics research?

• To do what humans cannot do

3
What is the point of robotics research?

• To do tasks that humans prefer not to do

From WALL-E
4
Current State in Home Robotics

• Often split in a myriad of subtasks navigation,
  recognition, scene understanding, manipulation,
  reasoning, etc.
• Boundaries and interfaces often ignored and are
  problematic
• Systems engineering is challenging ut generally
  not publishable
• Integrated systems are rare eg. Stanfords
  STAIR, etc.

5
Embodied Vision

• Actively seeing for some task
• How images are acquired are not considered
  traditionally in computer vision, encouraging
  unrealistic assumptions
• Eg. Benchmark datasets in object recognition
• not representative of actual situations
• learning algorithms rely on simplifications
• hard to evaluate whether systems work outside lab

6
What is SRVC?

• Photo scavenger hunt, where training data is
  acquired from internet

7
UBCs Experience

• Curious George (2007, 2008, )

8
UBC and Collaboration

• Integrated our lab
• New research directions
• Platform on which to test ideas
• Provides quick way to introduce new students

9
Designing a winner

• Good design choices
• Eye level camera on PTU
• Peripheral / foveal system with high res. camera
• Good Algorithms
• SLAM navigation
• Saliency and visual coverage
• SIFT based recognition
• Category recognition
• After initial phase, can now focus more on
  research

10
What does the SRVC do well?

• Compelling task
• Visibility
• AAAI 2007, Vancouver, Canada
• CVPR 2008, Anchorage, USA
• ISVC 2009, Las Vegas, USA
• Responsive to entrants
• Encourages open source
• Evolves
• Interesting for audience

11
Future Directions for SRVC

• Attract more competitors
• more synthesis
• greater exposure
• more exciting
• Improve research outcomes
• Research competitions should advance research
  rather than simply display current technology
• Should reflect successful research, not
  engineering that doesnt transfer

12
Attracting Competitors

• Currently
• 2 leagues, Software league and Robot league
• Software league is too similar to competitions
  like PASCAL VOC
• Robot league poses challenges due to shipping,
  unknown environment, etc.

13
Software League

• Offer more sensory modalities
• Stereo vision, high res images, video
• Offer mapping info, camera pose
• Larger test sets for more statistical validity
• Improve research outcomes (later)

14
Robot League

• Provide more detailed specifications for contest
  environment
• Provide standardized robot platform and
  architecture (like ROS)
• Avoids per team risk of shipping/unknowns
• Provides more opportunities for code sharing
• - Also involves numerous challenges
• Focus on more interesting challenges like
  viewpoint planning

15
Improving Research Outcomes
16
Improving Research Outcomes

• Improve realism more clutter, occlusions, no
  white tablecloths etc.
• Make context relevant
• Allow access to pre-built datasets and priors
• Web data is generally not suited for 3D
  recognition
• Forefront of vision research requires richer
  datasets
• Greater variety of objects and situations
• Points for

17
Conclusion

• Competitions can provide an evolving setting in
  which to evaluate current technologies
• SRVC frames a challenging problem for embodied
  vision, which is difficult to evaluate using
  benchmarks
• Numerous changes can be made to attract more
  competitors and improve research outcomes