Face Detection Service does not recognize biggest faces first

[calloatti]

Describe the bug

When submitting an image with 2 faces, when using "limit=1" the smallest of the 2 faces in the image is returned instead of the biggest face.

To Reproduce

Steps to reproduce the behavior:

1. Create a detection service
2. Send a request to detect faces in image using ""http://localhost:8000/api/v1/detection/detect?limit=0
3. 2 faces are recognized, smallest face first:

{
  "result" : [ {
    "age" : {
      "probability" : 1.0,
      "high" : 32,
      "low" : 32
    },
    "box" : {
      "probability" : 0.999154806137085,
      "x_max" : 106,
      "y_max" : 482,
      "x_min" : 3,
      "y_min" : 337
    }
  }, {
    "age" : {
      "probability" : 1.0,
      "high" : 38,
      "low" : 38
    },
    "box" : {
      "probability" : 0.9985826015472412,
      "x_max" : 360,
      "y_max" : 646,
      "x_min" : 153,
      "y_min" : 342
    }
  } ]
}

5. Send the same request using the same image but using "limit=1"
6. The smallest first face is returned instead of the second bigger face:

{
  "result" : [ {
    "age" : {
      "probability" : 1.0,
      "high" : 32,
      "low" : 32
    },
    "box" : {
      "probability" : 0.999154806137085,
      "x_max" : 106,
      "y_max" : 482,
      "x_min" : 3,
      "y_min" : 337
    }
  } ]
}

Expected behavior

When using "limit=1" on an image with 2 faces, the bigger face should be returned, not the smallest one, as described in the docs: "maximum number of faces on the image to be recognized. It recognizes the biggest faces first. Value of 0 represents no limit. Default value: 0"

Screenshots

Desktop (please complete the following information):

• OS: Windows 10
• Browser: Chrome
• Version: Latest

Additional context

Using Docker on Windows 10 subcenter-arcface-r100 custom build v 1.1.0

[calloatti]

Also, maybe add a "size" or "area" property to "box", with the value of the actual size of the face rectangle? Just to avoid having to do (x_max - x_min) * (y_max - y_min) for each face to get the value. I am calculating it now to avoid obvious false detections where the area is too small to be a real face.

[pospielov]

Thanks for reporting this. It looks like some models work properly, and some sort results by probability, not the face size. We will change it in the next release.