Automatic Generation of Robust Image Pair Correspondences for use in Sparse 3D Reconstruction

  • By @shailen@
  • 25 September, 2013
  • Comments Off on Automatic Generation of Robust Image Pair Correspondences for use in Sparse 3D Reconstruction
  • For any 3D reconstruction an accurate set of matchingsin a pair of images is required
  • Various methods suggested produce a large number of incorrect matchings
  • A study has been performed on the state-of-the-art matching generation algorithms to evaluate them objectively and propose modifications

Dataset

  • 24 images modeling
    • Translation
    • Rotation -Front parallel
    • Zoom
    • Affine Transformation
  • Five approaches to obtain correspondences tried over pairs in each category

Output

  • Development of a combination of techniques with some modification to ensure a high degree of correct correspondences

Technique 1 – Nearest Neighbours over SIFT

  • Problems
    • Unstable
  • Workaround
    • Try using SVD-matching using SIFT features

Technique 2 – SVD Matching using SIFT features

  • Technique
    • M features from image 1
    • N features from image 2
    • Construct a M X N matrix G
    • Perform SVD over G = U E VT
    • Replace Singular value matrix by Identity
    • P = U I VT
    • If P(i,j) is max in ithrow and jthcol
    • ith feature of image 1 matches jth feature of image 2

Technique 3 – PCA SIFT

  • PCA over SIFT keys
  • 128 D to 36 D keys
  • Found more reliable on our dataset

Technique 4 – SVD over PCA SIFT

  • Construct the matrix G using PCA-SIFT keys as in SVD-SIFT
  • Replace Singular value matrix by Identity
  • P = U I VT
  • If P(i,j) is max in ithrow and jthcol
  • ithfeature of image 1 matches jthfeature of image 2
  • Variations which were used
    • Different window sizes σ were tried
    • C(i,j)
      • correlation between ithand jthkeypoint
      • 0.5*(2-eucddist between kiand kj)

Technique 5 – Outlier Rejection

  • RANSAC over PCA
  • Found to be most reliable over our dataset

Major Comparisons are shown below

Plain Translation





Scale Variation (1)





Scale Variation (2)





General Affine (1)





General Affine (2)





 

Categories: Projects

Comments are closed.