Project 4 Artifact

John Gu
CSE455 Computer Vision
Project 4 Artifact
  • Testing Recognition with Cropped Class Images
  • Cropping and finding faces
  • Verify Face

    • Testing recognition with cropped class images

    Nonsmiling_cropped:

    Average face:

    Number of faces recognized vs. Number of eigenfaces used

    Questions
    The trend I see with the graph is that as the number of eigenfaces used increases, the number of faces recognized also increases.

    At 21 eigenfaces, the number of faces recognized maxes out. I would say that a good tradeoff is to use 5 eigenfaces since at 5 the number of faces recognized takes a substantial jump and then reaches a limit of the max number of faces, without having to do a lot of computing for the eigenfaces.

    Some bad matches by the program:




    The mistakes weren't too bad, the correct image was fairly high up with the matching.

    Cropping and finding faces

    elf.tga image, using min_scale = 0.45, max_scale = 0.55, step parameter = 0.01
    This resulted in just...the man's eye. A possible cause for this is because the eig10.face userbase that we were using is full of adults' faces, so it didn't recognize the baby's face.

    My "picture", using min_scale = 0.45, max_scale = 0.55, step parameter = 0.05
    This resulted in a cropped image of the glow of a lamp. This mistake may have occurred because the lighting pattern and the the way the glow covers the wall is similar to the gradience of a person's face. Also, the scale may not be small enough to shrink the image to detect all of a face.

    (Nonsmiling) IMG_0010.tga, using min_scale 0.45, max_scale 0.55, step parameter 0.05
    This recognized the faces correctly.


    (Smiling) IMG_0012.tga, using min_scale 0.45, max_scale 0.55, step parameter 0.05
    This was unable to pick up the face on the right (a false negative). An error that may have occurred is it may have sorted the matching faces incorrectly. I also accidentally had it search for 5 faces.

    IMG_0002.tga, using min_scale 0.45, max_scale 0.55, step parameter 0.05
    My face detection program seems to really like one particular student in the class. One mistake is my overlap function is incorrect, since it detects several faces that overlap each other. Another is the faces in this picture are harder to recognize since they are not from the users database.

    A group photo of four faces, using min_scale 0.45, max_scale 0.55, step parameter 0.05
    One reason that may account for this misrecognition of faces is that the overlap or sorting functions are incorrect, leading to hilarious face recognitions of feet.
     

    Verify Face

    I tried MSE thresholds of 500, 750, and 1000. A threshold level of 1000 worked the best, using verifyface as the search method.

    Using a MSE threshold of 1000, the false negative rate was 12.5%. The false positive rate was approximately 20%, my program was very good at choosing things that weren't faces to be faces.