Depth%20Edge%20Detection%20with%20Multi-Flash%20Imaging

1
Depth Edge Detection with Multi-Flash Imaging

• Gabriela Martínez
• Final Project Processamento de Imagem
• IMPA

2
Introduction

• Classic given a single two dimensional image,
  how can one detect edges of important features??
• Ramesh et al, introduce an algorithm based on
  multi-flash imaging, input 5 images.

3
Method

• The algorithm needs minimum of five images
  Ambient, and four flashes images positioned
  above, below, right and left of the lens.

4
Method

• To detect an edge passing trough a pixel,
  consider the epipolar ray corresponding to the
  line between the flash and the pre-image of the
  pixel.

5
Algorithm Description

• Ambient Image A
• n pictures with a light source Fk
• FkFk-A
• For all pixels x, Fmax(x)maxk(Fk(x))
• For each k create Rk(x)Fk(x)/Fmax(x)
• For each Rk traverse epipolar ray ek
• Find pixels y with negative transition, mark y

6
Remarks

• The value of ratio images at flash pixels is
  roughly1 for shadowed pixels, the value is
  close to 0.
• Intensity shows a sharp negative transition.

• Depth edge detection has been reduced to an
  intensity edge detection.
• It is easy to solve using Sobel kernel
  convolution.

7
Implementation

• The algorithm was implemented in matlab. To solve
  the intensity edge detection problem use Sobel
  kernel, generated by fspecial, and then use
  imfilter.
• Threshold After computing the confidence map,
  separate it in two images (low confidence 0.5,
  high confidence 1) then connect them using bwlabel

8
Comments

• The algorithm is easy to implement and it
  requires little computation.
• A robust classification to distinguish depth
  edges from texture edges.
• Making use of the epipolar relationship between
  flash and cast shadows to extract geometric
  features theres no need to create 3D scene
  reconstruction.

9
References

• Ramesh et al. Non-photorealistic Camera Depth
  Edge Detection and Stylized Rendering using
  Multi-Flash Imaging. ACM Siggraph 2004.
• Tien-Tsin Wong. Solving Visibility with Epipolar
  Geometry. The Chinese University of Hong Kong.
• Gonzalez R. Woods R. Digital Image Processing
  Using Matlab. Editorial Prentice Hall

10
Results

• Ambient Edges

11
Results

• Ambient Edges

12
Results

• Ambient Edges

13
Results

• Ambient Edges

14
Results

• Ambient Edges

15
Results

• Edges Color

16
Results

• Edges Color

17
Results

• Edges Color

18
Results

• Edges Color

19
Results

• Edges Color