Weekly Status Report

1
Weekly Status Report

• Edward Guldemond

2
Current Status

• It works!
• With cascade from last week, on various sets with
  default threshold of 0.5
• 100 detection on positive training images
• 0 detection on negative training images
• 79 detection on positive testing images
• 12 detection on negative testing images
• Results on training expected, results on testing
  good with 200 feature cascade

3
Current Status

• Lowering the threshold to 0.35
• 100 detection on positive training images
• 16 detection on negative training images
• 93 detection on positive testing images
• 31 detection on negative testing images

4
Current Status

• Raising the threshold to 0.65
• 100 detection on positive training images
• 0 detection on negative training images
• 34 detection on positive testing images
• 0 detection on negative testing images

5
Current Work

• Generated a new cascade of 200 features
• Expanded positive set to 98 images
• Expanded negative set to 2233 images
• From previous testing set added for robustness
• The testing set was the same as for the old
  cascade (possible problem)
• Only one positive eye image was taken - the one
  that was best aligned

6
Example of Images
7
Example of Images
8
Current Work

• New cascade rates at threshold 0.5
• 100 detection on positive training images
• 0 detection on negative training images

9
Current Work

• Lowering new cascade threshold to 0.35
• 100 detection on positive training images
• 24 detection on negative training images
• 93 detection on positive testing images
• 32 detection rate on negative testing images

10
Current Work

• Raising new cascade threshold to 0.65
• 100 detection on positive training images
• 0 detection on negative training images
• 34 detection on positive testing images
• 0 detection rate on negative testing images

11
Current Work

• The new cascade was run in scanning mode over a
  set of group photographs at differing thresholds
• Results are shown in the following slides

12
Results Part One
13
Results Part Two
14
Results Part Three
15
Results Part Four
16
Results Part Five
17
Results Part Six
18
Conclusions

• More training images are the way to go
• Correcting ratio of positive training images to
  negative training images
• Write code to auto-cut eyes based on output of
  current code