PANORAMA STITCHING

1
PANORAMA STITCHING

• Abhishek Agarwal
• Manas Agrawal
• Raghav Agrawal

2
Panorama Stitching

• Combining multiple photographic images with
  overlapping fields of view
• Producing a segmented panorama or high-resolution
  image

3
Approach

• Parameter Estimation
• Shift Estimation
• Projection Estimation
• Image Registration Transforming all the images
  to a common image plane.
• Blending Stitching the images
• compensate for exposure differences and other
  mis-alignments

4
Direct (pixel-based)

• Parameter Estimation Fourier Based shift
  estimation using phase correlation.
• Blending Average shift based

5
Feature Based Stitching
6
Feature Point Detector

• Various Approaches-
• Difference of Gauss (DoG)
• Bad Stability, Too many Scale Space Filter
• Speeded Up Robust Feature (SURF)
• Using the integral image to filter the image
  reduce the precision and stability
• Harris/Laplace
• Good stability, relatively inefficient, which is
  balanced by the reduction of scale samples
• Harris/Affine
• Too Complex

7
Feature Detection Harris
8
Harris Detector

• Change in appearance for shift u,v

9
Harris Detector - steps

• Compute Gaussian derivatives at each pixel
• Compute second moment matrix M in a
• Gaussian window around each pixel
• Compute corner response function R
• Threshold R
• Find local maxima of response function
• (non-maximum suppression)

10
Harris Detector - steps
11
Harris Detector - steps
12
Harris Detector - steps
13
Harris Detector steps
Take Local Maxima
14
Harris/Laplace

• Scale Adapted Harris Corner Detector
• Scale Adapted M

15
FEATURE MATCHINGSIFT(scale invariant feature
transform)

• SIFT is the most popular feature descriptor
  because of its stability and precision highly
  outperforms other descriptors.
• Each feature is assigned a descriptor vector and
  feature vectors are matched in different images.
• Advantage of SIFT
• Scale and rotation invariant
• Resistance of Noise in position and intensity
• High match precision

16
Implementation

• Get the 1616 Neighbor of every feature point.
• Find the gradient of every pixel in the
  Neighborhood.
• Build an orientation histogram for the feature
  point.
• Assign the feature point a principal
  orientation.
• The gradient of every pixel in the Neighborhood
  is rotated according to the principal orientation
  (for rotation invariance)
• Split the 1616 Neighborhood into 44 squares,
  calculate the orientation distribution in each
  square, and build a descriptor.
• Normalize the descriptor vector

17
Feature Matching

• Best match is defined as the minimum of the sum
  of the absolute difference between descriptor
  vectors.
• There are several speeded up matching method.

18
PARAMETER ESTIMATIONTransformation Matrix
Estimation

• Transformation Model
• Principle (RANSAC)

19
Blending

• Gain Compensation
• Since the transform matrix is known, the two
  image can be simply aligned together.
• The stitched image here has obvious artifact.
  The most direct one is the difference of
  intensity.
• Gain Compensation is a fundamental way to
  eliminate intensity difference.

20
Implementation (Gain Compensation)

• Find the overlap region of the two aligned image.
• Calculate the average intensity of the overlap
  region of the two image respectively.
• Calculate the difference of the average
  intensity of the overlapped region.
• Add the difference of the average intensity to
  one of the image wholly, Then ,the two image
  average intensity in the overlapped region should
  be the same.

21
References

• R. Szeliski. Image alignment and stitching A
  tutorial. Technical Report MSR-TR-2004-92,
  Microsoft Research, Last updated, December 10,
  2006.
• D. Lowe. Distinctive image features from
  scale-invariant keypoints. International Journal
  of Computer Vision, 60(2)91C110, 2004.
• M. Fischler and R. Bolles. Random sample
  consensus A paradigm for model fitting with
  application to image analysis and automated
  cartography. Communications of the ACM,
  24381C395, 1981.
• M. Brown and D. G. Lowe. Automatic panoramic
  image stitching using invariant features. Int. J.
  of Computer Vision, 74(1)59C73, 2006.
• Brown, M. Lowe, D.G.. Recognising panoramas.
  Computer Vision, 2003. Proceedings. Ninth IEEE
  International Conference on. 13-16 Oct. 2003
  Page(s)1218 - 1225 vol.2

22
QUESTIONS??