Ensure correctness of faceness measure #56
Comments
jakewilliami
added a commit
that referenced
this issue
Nov 4, 2021
|
The difference in faceness scores between positive and negative images seems to be statistically significant after some changes, just using 1000 training images and 1000 testing images: |
jakewilliami
added a commit
that referenced
this issue
Nov 9, 2021
This was previously incorrectly calculated. The results for this are looking much more promising now. The method of calculation is more in-line with what was originally proposed: > As for your question about adapting the Viola-Jones face detector to return a measure of face-likeness - I think one simple thing you could use is the number of features computed before terminating from the cascade. Face patches always make it to the end of the cascade, i.e. all Haar-like features are computed. Non-faces usually result in early termination from the cascade, and the earlier, the less face-like the patch is (roughly speaking). This certainly would not be a perfect measure, but I'm confident that there is a positive correlation between the total number of features computed in the cascade and the "face-likeness".
jakewilliami
changed the title
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
This is an adapted feature. I think I slightly misunderstood, and this needs to be corrected.The faceness measure needs to be reviewed to ensure its correctness. It should be a feature of the cascade, which itself needs to be assessed.
The text was updated successfully, but these errors were encountered: