Depth ordering

1
Depth ordering

• Guimei Zhang
• MESA (Mechatronics, Embedded Systems and
  Automation)LAB
• School of Engineering,
• University of California, Merced
• E guimei.zh_at_163.com Phone[phone]
• Lab CAS Eng 820 (T 228-4398)

Sep 22, 2014. Monday 400-600 PM Applied
Fractional Calculus Workshop Series _at_ MESA Lab _at_
UCMerced
2
Introduction

• What is depth ordering?

(a) Imput image
(b) Edge image
(b) Depth ordering
3
Introduction


(a) Imput image
(b) Edge image

(c) Contour completion
(d) Image layer
4
Introduction

• Applications (why to do this work?)
• Image segmentation
• Object recognition
• Target tracking
• Scene understanding

5
Depth ordering algorithm based on T-junctions
and occlusion reasoning

1. Motivation
2. Method
3. Experiments
4. Conclusion

6
1. Motivation

• T-junction points
• convexity

7
1. Motivation

Problems

• Existed methods have limitations to order objects
  completely, especially in multiple backgrounds .

8
1. Motivation

• conventional methods always detect T-junctions
  before segmentation, which will result in
  detecting false T-junctions or missing real
  T-junctions in clutter images

9
2. Method

• overcomes the first problem by introducing high
  level occlusion reasoning theory when some
  regions include no T-junction, no convexity or
  inconsistent T-junction point

10
2. Method

• We combine low level depth cue (T-junctions) and
  high level occlusion reasoning, therefore make
  progress to order the objects completely, even in
  multiple backgrounds.
• In addition, conventional methods always detect
  T-junctions before segmentation, which will
  result in detecting false T-junction or missing
  real T-junctions in clutter images.

11
2. Method
2.1 T-junction analysis

• Character 1 T-junction is composed by three
  boundaries and only two boundaries are collinear,
  in other words, the angle between them is 180
  degree. Two collinear ones are named as occlusion
  boundaries, and the other is called occluded
  boundary.
• Character 2 The region contained occlusion
  boundaries is in front of the one included
  occluded boundary.

12
2. Method

• In previous work, T-junctions are detected
  before segmentation. The shortcomings of this
  kind of methods are as follows
• it is easy to detect false T-junctions due to the
  complexity of the real images and texture of some
  objects

13
2. Method

• preserve T-junctions before image segmentation
  and remove the false T-junctions, in other words
  the post-processing is time consuming
• detection T-junction method based on image is
  more complex than one based on contour. So we
  first segment real image and get the contour of
  image, then detect T-junctions on the contour
  image.

14
2. Method

• Detection T-junction points

09/22/2014
AFC Workshop Series _at_ MESALAB _at_ UCMerced
15
2. Mehtod
2.2 occlusion reasoning

• visual psychology principle
• The figure (foreground) has definition shape, but
  the background has not, if the background is
  perceived as having certain shape, that is due to
  the other gestalt.
• The background seems continuous stretch without
  being interrupted behind the figure.
• The figure always appears in the front and the
  background is in the back.
• The figure can give human more deep impression,
  and easier to remember.

16
2. Mehtod

• Reasoning laws( inspired by human cognition)
• Law 1 If the background has not definition
  shape, the region which has definition shape is
  in front of the one which has not.
• Law 2 When the background has definition shape,
  we first remove part objects formed the boundary
  of background, and can get the region which has
  definition shape is in front of the one which has
  not.

17
2. Mehtod

• Law 3 The lower the background region in the
  image is more likely to be closer to viewpoint
  when there are multiple background regions in the
  scene.

18
2. Method

• the method is as follows

19
3. Experiments
Experiment result
First input image
Sec T-junction detection
Last The depth map
( rendered as a gray level image, and high values
indicate regions closer to the viewpoint)
20
3. Experiments
Experimental results
21
3. Experiments

• Comparison with the state of the art

(a) input image (b) T-junction detection (c)
The depth-map obtained by the method in Ref 7
(d) The depth-map obtained by our method
22
3. Experiments
(a) input image (b) segmentation (c) The
depth-map obtained by the method in Ref 8 (d)
The depth-map obtained by our method
23
3. Experiments
(a) input image (b) T-junction detection of our
method (c) The depth-map obtained by our method
(d) T-junction detection of Ref 6 (e) The
depth-map obtained by the method in Ref 6
09/22/2014
AFC Workshop Series _at_ MESALAB _at_ UCMerced
24
(No Transcript)
25
4. Conclusion

• A new T-junctions detection method based on
  contour is proposed in this paper, which can
  accurately detect the T-junctions on an already
  segmented image.
• And Monocular depth ordering algorithm based on
  low level depth cue (T-junctions) and high level
  occlusion reasoning is proposed in this paper.

09/22/2014
AFC Workshop Series _at_ MESALAB _at_ UCMerced
26
4. Conclusion

• The initial depth image ordering is first
  obtained based on T-junction and then more
  detail depth ordering can be achieved by using of
  high level occlusion reasoning.
• Results are compared with the method using depth
  cue (T-junction and convexity) and the method
  optimization algorithm based frameworks, our
  method can get the perfect depth ordering, and
  can establish global and consistent depth
  interpretation.

09/22/2014
AFC Workshop Series _at_ MESALAB _at_ UCMerced
27

• Thanks

AFC Workshop Series _at_ MESALAB _at_ UCMerced
09/22/2014