An opposition to: Context-Based Vision System for Place and Object Recognition Contextual Models for Object Detection Using BRFs

1
An opposition to Context-Based
Vision System for Place and Object Recognition
Contextual Models for Object Detection
Using BRFs

• Authors Antonio Torralba, Kevin P. Murphy,
  William T. Freeman, and Mark A. Rubin
• Opponent Carlos Vallespi

2
Paper claims

• Claims to recognize 63 different locations.
• Claims to categorize new environments
• Claims to help object recognition by suggesting
  presence and location.

3
Place recognition
Is the classifier really doing anything?

• Temporal information is available.
• HMM will help a lot to the classifier.
• Only 2-3 choices are possible at a time, knowing
  the current state.

4
Simple place recognition with SIFT
Database
5
Simple place recognition with SIFT
Test DB
6
Comparing with SIFT
74 matches
7
Comparing with SIFT
Some correct matches
8
Comparing with SIFT
Correct no matches
9
Comparing with SIFT

• No incorrect mismatches
• Just one weak match (22 matches)
• Provided 9 locations and 100 accuracy in the
  test set.

10
Scene categorization

• This paper claims that they are able to
  categorize 17 unseen scenarios.
• We have seen other methods in the past for scene
  categorization that also worked well (with up to
  13 classes)
• Bag-of-words approaches (using textons, for
  instance).
• Histogram-based approaches.
• Torralbas paper (using image frequencies).
• They use an average of local features over the
  image with a sliding window.
• In fact, this is just a sort of histogram
  approach (nothing new).
• DB does not seem very generic. They do not
  compare with other methods.
• It performs poorly, except for the exception of
  the HMM

11
Object presence and location

• Their own images speak for themselves )

• A filecabinet is expected to be seen in almost
  the entire image.
• Most of the objects that are highly expected to
  be found, do not show up.

12
Object presence and location

• Their own images speak for themselves )

• Except for the case of the building (which I am
  sure I could get something similar by averaging
  all the bounding boxes of buildings), all others
  are wrong even the sky.

13
Conclusions

• Place recognition
• It seems to be an easy problem, that can be
  solved by simpler methods without temporal
  information.
• An HMM alone could have done similar work.
• Scene categorization
• Suspicious DB
• Only works because of the temporal information.
• Object presence and location
• Just does not work.