Face counting inaccurate with latest firmware

OpenMV Boards
1

My H7 Plus was just updated to firmware 4.5.5, and code that previously worked quite accurately is now showing face counts of 2 when there is only a single face in the image. Key parts of the code below.

# Reset sensor
sensor.reset()
sensor.set_contrast(3)
sensor.set_gainceiling(16)

# HQVGA and GRAYSCALE are the best for face tracking.
sensor.set_framesize(sensor.HQVGA)
sensor.set_pixformat(sensor.GRAYSCALE)
sensor.set_framebuffers(1)

sensor.skip_frames(time = 2000)     # Wait for settings take effect.

# Load Haar Cascade
# By default this will use all stages, lower satges is faster but less accurate.

face_cascade = image.HaarCascade("frontalface", stages=25)
print(face_cascade)


faceCount = 0;

# The set of angles to rotate the raw image and determine the best orientation for face detection

angles = [-45, 0, 45]

rimg = sensor.alloc_extra_fb(240,160,sensor.GRAYSCALE)

totalFaceCount = 0

while (True):
    # Capture snapshot

    img = sensor.snapshot()
    img.gamma(gamma=1.0,contrast=1.5,brightness=0.0)

    maxFaceCount = 0

    # Account for tilted faces or a tilted camera

    for angle in angles:
        rimg.replace(img)
        rimg.rotation_corr(x_rotation=0.0, y_rotation=0.0, z_rotation=angle)

        # Find objects.

        objects = rimg.find_features(face_cascade, threshold=0.75, scale_factor=1.25)

        newFaceCount = len(objects)

        if(faceCount != newFaceCount)
               faceCount = newFaceCount
               print(faceCount)
2

Hi Rick,

We haven’t changed anything with the functions you are mentioning recently.

It is not the case though that behavior may stay exactly the same if we just something or improve a feature.

Can you determine where the script is having issues? When does it detect extra faces from what it did previously?

…

I can test your code in about 12 hours and report my own results.

3

Hi! I’m as surprised as you are. The code would (correctly) always return either a zero (0) or one (1) value when running it - and just recently after the firmware update it started returning two (2) intermittently, even with a full frontal face image with a single face (mine!) in it. Let me know if there’s any suggestions for code changes or configuration/settings and I’ll be happy to test on my end also. Thank you!

4

Hi Rick,

I ran the code on my OpenMV Cam H7 Plus with the latest firmware and I just get 0/1 printed alternating over and over again. You are probably experiencing a different issue.

import sensor
import image

# Reset sensor
sensor.reset()
sensor.set_contrast(3)
sensor.set_gainceiling(16)

# HQVGA and GRAYSCALE are the best for face tracking.
sensor.set_framesize(sensor.HQVGA)
sensor.set_pixformat(sensor.GRAYSCALE)
sensor.set_framebuffers(1)

sensor.skip_frames(time = 2000)     # Wait for settings take effect.

# Load Haar Cascade
# By default this will use all stages, lower satges is faster but less accurate.

face_cascade = image.HaarCascade("frontalface", stages=25)
print(face_cascade)


faceCount = 0;

# The set of angles to rotate the raw image and determine the best orientation for face detection

angles = [-45, 0, 45]

rimg = sensor.alloc_extra_fb(240,160,sensor.GRAYSCALE)

totalFaceCount = 0

while (True):
    # Capture snapshot

    img = sensor.snapshot()
    img.gamma(gamma=1.0,contrast=1.5,brightness=0.0)

    maxFaceCount = 0

    # Account for tilted faces or a tilted camera

    for angle in angles:
        rimg.replace(img)
        rimg.rotation_corr(x_rotation=0.0, y_rotation=0.0, z_rotation=angle)

        # Find objects.

        objects = rimg.find_features(face_cascade, threshold=0.75, scale_factor=1.25)

        newFaceCount = len(objects)

        if(faceCount != newFaceCount):
               faceCount = newFaceCount
               print(faceCount)