Identifying small faces in large images

[ashlinghosh]

I'm trying to retrain retinaface with a custom dataset. My images are of resolution 1920x1080. The average width and height of faces in the images are ~20 pixels. I have around 10k images for training. So far, the model is not able to identify faces. Is there any preprocessing i can do like resizing and cropping which can help in improving the detection accuracy?

[Counterfeiter]

I try the same... If you haven't infinity video memory the only solution is to crop images with overlapping sections (to don't miss a face). This ends in doubled detected face:

def cropwithoverlay(img, max_w, max_h, overlap_per = 25.0):
    stepsize_w = int(float(max_w) * (100.0 - overlap_per) / 100.0)
    stepsize_h = int(float(max_h) * (100.0 - overlap_per) / 100.0)
    rows, columns, _ = img.shape
    img_crop_list = []
    if rows <= max_h and columns <= max_w:
        return [[0,max_h,0,max_w]]
    for row in range(0, rows, stepsize_w):
        row2 = min(row + max_h, rows)
        for col in range(0, columns, stepsize_h):
            col2 = min(col + max_w, columns)
            img_crop_list.append([row,row2,col,col2])

    return img_crop_list

To calculate the boxes in the large image... My current try is to detect min 50 % overlapping boxes of the largest box and mark this compare as doubled detected box...

def faceboxisdoubled(a, b, threshold = 0.49):  # returns False if rectangles don't intersect
    dx = min(a[2], b[2]) - max(a[0], b[0])
    dy = min(a[3], b[3]) - max(a[1], b[1])
    #intersect?
    if (dx>=0) and (dy>=0):
        max_box_sqr = max((a[2] - a[0]) * (a[3] - a[1]), (b[2] - b[0]) * (b[3] - b[1]))
        if (dx*dy) > (max_box_sqr * threshold):
            return True

    return False

49 % is sometimes possible and valid in large crowds... you could play with this value...

I stuck at a function to delete the double (or tripple) detected boxes from the list and keep the larger one. I should try to sleep more :D. If someone has the solution please write it down. Thx

best regards

Basti

[ahmetybilgin]

@Counterfeiter Have you continued to implement the solution?