Shape From Silhouette

1
Shape From Silhouette

• Chris Boehnen
• Fall 2003
• 3D Photography

2
Outline

• How/Why does it Work?
• Limitations
• Advantages
• My Implementation
• Results
• Conclusions

3
How SFS Works

• SFS Extracts the Silhouette from an object and
  using the known camera orientation and multiple
  images. Using this it is able to carve away
  sections of space which the object could not
  exist in.
• The remaining sections of space are a visual hull
  of what the object is.

4
How SFS Works
5
How SFS Works
6
Limitations

• This means SFS is not able to capture concavities
  in a surface with respect to the current view

Top View of Object
Camera View
Top View of Object
Camera View
Area is red is mis-calculated due to Concavity
from camera view
7
Limitations

• Silhouette extractions is commonly done using a
  background of fixed color which means that the
  object being scanned can not contain that color.

If the background color is black, and the eyes
and mouth of the smiley face is black, the
algorithm assumes the eyes are holes leading
through to the background
8
Advantages

• Low Cost hardware, only a camera and a turn table
• While typically fairly low resolution, models can
  be texture mapped nicely to help compensate
• Captures watertight models of the entire object

9
My Implementation

• In order to simplify the calibration process, the
  camera and turn table must be level. This
  ensures that the axis of rotation represents the
  same line in the 2D Image for every image taken
• Intrinsic properties of the camera are calibrated
  out using Bouguets Matlab code.
• SFS is then calibrated automatically based on a
  few input factors determined solely from the
  images themselves

10
Calibrating the Axis of Rotation

• Since we wanted to be able to calibrate based
  solely on the input images, finding the axis of
  rotation is difficult. However, with our
  constraints we only need to know where it lies in
  the image plane because it will then lie in the
  same location for everything.

11
My Implementation

• Images are captured using a turntable and a
  Cannon Powershot G2 Digital Camera. Acquisition
  of the sequences of images is automatic and uses
  a piece of software written specifically for this
  purpose.

12
Calibration Procedure, Matlab Code

• 1. Locate the corners in the calibration
  images.
• 2. Execute the calibration program to obtain
  intrinsic parameters.
• Code Calculates Focal Length in Pixels, and the
  Principal Point which are used later
• 3. Apply intrinsic parameters to each input
  images and resave them once they have been
  rectified.

13
Calibration Procedure, Matlab Code

• Extreme corners of the checkerboard pattern are
  located and clicked on.
• Program automatically detects and extracts each
  corner of the checker board pattern.

14
Calibration Procedure Matlab, Code

• Program computes intrinsic/extrinsic parameters
  of the camera.
• Parameters are then used to remove distortion of
  each input image.

15
My Implementation

• The images are then separated into background and
  object and re-saved so they can be viewed. The
  images could be manually modified at this stage
  if a color of the object matched the background
  color to eliminate the eyes on the smiley face
  etc. Small holes less than 10 pixels are filled
  automatically

16
Top View of Projection Process
Image Plane
Principal Point
Focal Length
Volume to Be Carved Origin
Focal Point (Cos F, Sin F)
Axis of Rotation Line
17
Top View of Projection Process
Each valid voxel is projected onto the image
plane. Then, it is determined if its location
on the image plane is an object or background
location

18
Implementation

• This method is repeated for each image. Only
  valid voxels are projected because invalid ones
  can already be thrown away. By not worrying
  about voxels already marked invalid the process
  runs significantly faster.

19
Converting Volumetric data to Polygonal data

• What we now have is volumetric data, but are
  interested in the surface data. So, what we do
  is only draw polygons on points on the outside
  surface.
• So you loop over each valid voxel. And if the
  voxel to the left of it is invalid, you draw the
  left hand side of it, and repeat for each side

20
Demo and Results Time