Stereo: Cyclopean surround interactions, stereomatching model and beyond.

1
Stereo Cyclopean surround interactions,
stereo-matching model and beyond.

• Do cyclopean boundaries pop-out?
• Disparity capture as a measure for cooperative
  interaction in stereopsis
• A model of stereo-matching in V1 and V2
  (preliminary results)

2
Do exclusively binocular (cyclopean) boundaries
pop-out?

• Boundaries in natural images are usually visible
  to either eye separately. Then they are known to
  pop-out, i. e. are easily detectable. What if
  boundaries become visible only after the
  monocular images are fused together into a single
  3D image?
• Are they going to
• pop-out?

3
(No Transcript)
4
(No Transcript)
5
(No Transcript)
6
Conclusions

• Unlike monocular, cyclopean boundaries are only
  easily detectable when they are defined by
  luminance magnitude and sometimes motion
  contrast.
• Color, luminance polarity, and orientation
  defined boundaries do not pop-out.
• Bakin and Nakayama showed that unlike V2,
  contextual interaction in V1 are not
  depth-sensitive.
• This suggests that the functionality of
  contextual interactions in V2 is different from
  V1.

7
Disparity capture as a measure for the
cooperative interaction in stereopsis

• Cooperative interaction is the backbone element
  in few important stereopsis algorithms (Marr
  Poggio I, PMF, Prazdny).
• Establish if the cooperative interaction is in
  effect.
• Find the range of the interaction and its
  scaling properties.

8
(No Transcript)
9
(No Transcript)
10
Conclusions

• A short-range cooperative interaction has been
  observed in a simple stereoscopic matching task.
• The range of the observed interaction scales
  proportionally to the stimulus size and varies
  between different subjects from 1.5 to 3 stimulus
  sizes.
• The results provide some support for the
  cooperative interaction, but the short range of
  the observed interaction makes it impractical as
  a stand-alone stereoscopic mechanism.

11
Stereo Model

• Based on neuron types existing in V1 and V2.
• Solves correspondence problem by calculating a
  local correlation measure between left and right
  eye images.
• The proposed correlation measure approximates a
  scalar product between the matched images via a
  simple network of simple and complex cells.
• Surround interactions can be added in a
  straightforward way, which can be used for
  boundary detection and illusory contour formation
  (future work).

12
(No Transcript)
13
Flat neuron
Tuned excitatory neuron
Left eye
Left eye
Right eye
Right eye
14
(No Transcript)
15
V2
the same depth
facilitation
V1 binocular
V1 monocular
phase
summation
LGN
16
(No Transcript)
17
(No Transcript)
18
(No Transcript)
19
(No Transcript)
20
Conclusions

• A biologically plausible model of stereo
  processing in V1 and V2 is proposed.
• The task of stereo-matching is carried out via a
  network of connections reaching from LGN to V2,
  which calculates a correlation measure between
  left and right eye images. The measure
  approximates a scalar product between disparate
  local regions for these images.
• Preliminary results show that the model solves
  the stereo-matching task for a wide range of
  inputs, including transparent and left/right eye
  unbalanced stereograms.