1 of 13

Classical Approach Results

2 of 13

How SSIM Works

  • MSE finds absolute errors between individual pixels
  • SSIM finds changes in structure of image (inter-dependencies between pixels)
  • Output:
    • -1: completely different
    • 1: exactly same

3 of 13

How SSIM Works

x and y are NxN patches of the two images

4 of 13

How SSIM Works

Luminance: if mu_x == mu_y, l(x,y) = 1

​

​

Contrast: if sigma_x << sigma_y, c = 0

​

Structure (c3 = c2/2): Var[XY] = Var[x] * Var[y] if

5 of 13

Sources

https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=1292216

https://en.wikipedia.org/wiki/Structural_similarity#cite_note-:0-4

​

6 of 13

Steps Taken:

  • Generating difference images with sparse data
    • only sparse negative and positive images taken
    • need direct matching positive-to-negative pairs to generate difference image
    • took all unblocked negatives into separate folder, ran algorithm to find closest timestamp matching positive to negative
  • Ring Detection
    • Tuned RGB thresholds for each of the three rings
    • In cases where one or two rings are missing, approximate their radius using the known rings
  • Scoring
    • Check from order of inner to outermost ring which one the axe tip lay inside using radius of ring and distance from tip to center

7 of 13

Results

what this image shows

8 of 13

Results

9 of 13

Results

10 of 13

Source

Detected Axe and Rings

11 of 13

Overall

12 of 13

Normalized

13 of 13

Conclusions + next steps

  • how this currenet method works
  • nexr steps with segmentation + classification