Shape Descriptors for Maximally Stable Extremal Regions by 8ubrKB

VIEWS: 36 PAGES: 17

									Shape Descriptors for Maximally Stable
         Extremal Regions


       Per-Erik Forss´en and David G. Lowe
       Department of Computer Science
       University of British Columbia

       Eleventh IEEE International Conference on
       Computer Vision(ICCV – 2007)
             Reporter:Shih-Hao Wang
                                                   1
Outline

   Introduction

   Multi-Resolution

   Maximally Stable Extremal Regions(MSER)

   Experiment Model

   Experiment

   Conclusions
                                              2
Introduction

   Affine-invariance Concept:originally features of images
    will not be changed after affine transformation

   Problem:When a scene is blurred or viewed from
    increasing distances, many details in the image disappear
    and different region boundaries are formed

   Use An affine invariant shape descriptor for maximally
    stable extremal regions (MSER) to reduce above-
    mentioned effect


                                                                3
Multi-Resolution

   Nested Vector Space:
V       V1  V0  V1          V




                  V3 V2 V1 V0    V1    V2   V3




                                                    4
Multi-Resolution




                   5
Maximally Stable Extremal Regions

   Apply a series of thresholds – one for each grayscale level.

   Threshold the image at each level to create a series of
    black and white images.

   One extreme will be all white, the other all black. In
    between, blobs grow and merge.




                                                               6
Maximally Stable Extremal Regions

   Example




                                    7
Multi-Resolution MSER

   Method:

   Step1:Instead of detecting features (from step4) only in
    the input image

   Step2:Construct a scale pyramid with one octave
    between scales

   Step3:Detect MSERs separately at each resolution

   Step4:Duplicate MSERs are removed by eliminating fine
    scale MSERs with similar locations and sizes as MSERs      8
    detected at the next coarser scale
Multi-Resolution MSER

   Scale pyramid




   The scale pyramid is constructed by blurring and resample
    with a Gaussian kernel ,  = 1.0 pixels.
                                                   x2  y 2
                                    1            
               G2 D ( x, y; )              e      2 2
                                   2   2                    9
Experiment (performance Measurement)

   Performance measure in this paper:inlier frequency

   k-th tentative correspondence is an inlier:inlier function
    is 1

   Otherwise is 0




                                                                 10
Experiment

   3D scene correspondence evaluation




   The 56 accepted correspondences are shown in blue.
    Features from rejected correspondences are shown in
    green.
                                                      11
Experiment

   3D scene results




   Performance of shape and texture descriptors on
    the scene
                                                      12
Experiment
   Planar and parallax-free scenes




   A scene with de-focus blur. The multi-resolution MSER
    provides better performance than using only the original   13

    resolution
Experiment




   Top image shows the 22 correspondences found using
    single resolution MSERs.

   Bottom image shows the 53 correspondences found using
    multi-resolution MSERs
                                                       14
Experiment
   Planar and parallax-free scenes




   A scene with scale change. Again, the multi-resolution
                                                         15

    MSER gives better performance
Experiment




   Top image shows the 22 correspondences found using
    single resolution MSERs.

   Bottom image shows the 33 correspondences found using
    multi-resolution MSERs
                                                       16
Conclusions

   New method provide better robustness to large scale
    changes and blurred images

   Improve matching performance over large scale changes
    and for blurred images




                                                       17

								
To top