Personal Site

 

 

Implementation

1. Feature points detection and description: MOSP

2. RANSAC

3. Blending method: linear blending

 

 

 

 

 

I. Take pictures

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
     

 

We took a serious of 17 photo in the square of National Chiang Kai-shek Memorial Hall. Note that the Moiré effects or aliasing in the ground is duo to the downsampling, there are no such pattern in the original size images.

 

II. Warp to cylindrical coordinate

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
     

 

In the second step, we warped all the photos using 1150 as the focal length of this case, and the warping center was the middle point of the horizon.

 

III. Feature detection and description: MOPS

We use MOPS (Multi-Scale Oriented Patches) algorithm.

Feature Detection:

  1. We work with grayscale images I(x, y). We form an image pyramid:

     

And we set s = 2, σp = 1.0.

  1. For each level of pyramid, we compute its Harris corner detector:

  

σi = 1.5, σd = 1.0.
Corner detection function:

And we choose point that is local maximum of 3×3 grid and larger than 10.

  1. We hope that the feature points that are generated are spatially well distributed over the image, so we use ANMS (adaptive non-maximal suppression) method:

At first we set radius R as infinity, and for each feature point, if it is the local maximum in this radius, then we pick this point.
Then we decrease R to repeat this process the number of feature points is satisfied.

  1. For each feature point, we compute sub-pixel refinement:

  

Feature Description:

  1. Compute orientation of feature point:

  

  1. At this orientation, we construct an 8×8 square as our descriptor. We rotate this square to θ= 0, and we sample from

 with spacing = 5 (choose central point of 5×5 patch as one descriptor).

  1. Normalize this 64-dim descriptor to [-1, 1]

 

VI. Compute pariwise alignments: RANSAC

 

V. Alignment & Blending

those translations that were computed by previous step to align the images.

.

 

 

VI. Result

 

. ----------------------------------------------------------------------------------

 

 

VII. Reference

1. Matthew Brown, Richard Szeliski, and Simon Winder, Multi-Image Matching using Multi-Scale Oriented Patches, CVPR 2005.