Annotating Images of Faces with ImageStats

Facial Annotation Tasks

This page last updated Jan 11th.

Deadline has passed to claim ImageStats tasks
As previously announced, 23:00 UTC Friday, January 11, 2013, was the deadline. Unclaimed tasks have been removed.

ImageStats was updated Friday, Jan. 4. Changes in features or procedures are highlighted by - NEW - below.

Displaying your Images

With ImageStats open, in the left-hand side, there is the “Event List”, which should have only your specific task, e.g. Face 01_31.

- NEW - To the right, in the “Search” results, you will see a list of images with anonymized filenames.

(In previous configurations of ImageStats, you had to first put a checkmark in your task's box, then hit “Search” [without altering what’s in the Search text field, “:”.])

Above them is a row of “Export” related controls. You will not need to use these except when you’ve completed your task.

Handling Each Picture - Key Steps

  1. Open picture
  2. Make Annotation Changes, If Any
  3. Either:
    • “Save” & “Close”
    • “Save” & “Next” (recommended) - NEW -

Handling Each Picture - Details

1) Open Picture

To open a picture, either:

  • double-click on it from a Search results page
  • - NEW - navigate to it from a previously open picture using the “Next” or “Previous” button. These buttons move in sequence through all the pictures shown within the current page of the Search results. In the case of “Next”, once you come to the last image in the current page, the button will be disabled, and you must use Save & Close instead, then change the Search results page. Similarly for “Previous”.

Once opened, the image should already have annotation rectangles (aka “boxes”) atop it, from a computer “FaceMatch” algorithm. These boxes may take a few moments to appear. (If you are revisiting an image you annotated earlier, see “Revisiting a Photo - Why Don't I See My Annotations?” below.)

- NEW - The radio buttons like “face” and “profile” now control not just the drawing choice, but what type of boxes you see. To see both frontal face and profile boxes at the same time, you must select “show all”. Editing is not possible in “show all” mode.

2) Make Annotation Changes, If Any

The “FaceMatch” algorithm provides boxes that can be wrong. Among your tasks:

  • Leave a box as is, if it’s good enough;
  • Change the size and/or location of boxes;
  • Delete unnecessary boxes;
  • Add missing boxes;
  • Change the type of a box from “face” (in yellow) to “profile” (in red), or vice versa.

Look at the picture carefully. For each face within it where features can be clearly seen, decide if it’s a frontal facial image (called simply “face”) or a “profile” facial image. If there’s no existing box of the right type, select the appropriate radio button, and then draw the bounding box around the face.

For details of how to draw or delete bounding boxes, see

- NEW - As mentioned above, you must be in either “face” or “profile” mode to both see and manipulate the corresponding type of box. “Show all” mode does not allow manipulation.

- NEW - Boxes for Face tasks (but not Skin tasks) will have a minimum height and width imposed. As a side-effect, it will no longer be possible to inappropriately “delete” a box by reducing a dimension to zero.

- NEW - You are not allowed to draw any box edge outside the picture boundary. (This solves a bug affecting “Save” in previous versions.)

Is it a Face or a Profile? Abstractly, this transition occurs when the face is turned about 45 degrees from the viewer. More concretely, if you can see both eyes including their outside edges (e.g., parts of eyelids furthest from nose), probably call it a “face”, otherwise a “profile”. If there is hair of something else obscuring your view of the eyes, you may have to infer the head angle. If in doubt, prefer “face”.

- NEW - If the type is wrong, delete the box (as described in Help), change the drawing mode to the correct type, and create a new box.

Where Should I Draw the Box from Scratch? Ideally, the box should be fairly tight around the main facial features, and not include much hair from the head. If the face is upright and not tilted, then, approximately,

  • the top of the box should be in the vertical middle of the forehead, above the eyebrows. That is, halfway between the eyebrows and the hairline (or where it would be if no balding).
  • the bottom of the box should be midway between the chin-point and bottom of the lower lip
  • for a profile, the box side near the nose should just be touching or a pixel or two beyond the entire head. For the box side near the back of head, the box should usually include the visible cheek and ear, but minimize the inclusion of background behind the back of the head.
  • for a frontal face, the box sides should be near the base of the visible ears, but always including nose, cheeks, and other main facial features.

If the face is tilted or not upright, the box may need to enlarged or otherwise adjusted so that main facial features are fully included.

If the Algorithm Draws the Box, When is a Side Good Enough to Leave Unaltered? Redraw a side that is much more tight (tends to clip features) than you would draw yourself. As for the the looser/larger direction, you should alter a side of the box that is more than a “finger's width” beyond the top or sides of a head or facial feature, or more than a “finger's width” below the chin down the neck. (“Finger's width” referring to an imagined finger placed against the person's physical head, if you were living in the scene shown.)

Height and Width of a Box. Whenever possible, avoid drawing or allowing boxes where the height is more than twice the width, or vice versa.

- NEW - Overlapping Boxes. Overlap between annotations is permissible and sometimes necessary, but try to avoid it or minimize its extent as much as possible.

Example Photos

Boxes determined by the algorithm include some unnecessary ones (left); these are deleted, but remainder are judged to be OK in both position and type (right):

Initial FaceMatch annotations  Corrected annotations

A typical frontal “face” box (left) and “profiles” boxes (right):

Frontal face example  Profile example

About these Photos. There is always a certain amount of subjective judgement in annotation. For the face box in the lower row above, the annotator chose to draw or allow a box that:

  • includes most of the ear seen on the left (while not having too much of the background); and
  • excludes most of the ear seen on the right (more in line with the guidance to clip at the base of the ear).

Both sides are within the acceptable range. When it doubt, it is desirable to minimize the background, because it can be mislead the future training of the face-finding algorithm. So the ear on the right is closer to optimal.

In the profiles example, the right side of each box is at the rear edge of the ear. This is consistent with the (updated Dec 18th) guidance, not to extend the box to or beyond the back of the head. We are interested in the face, not in hair or other parts of the image.

Incomplete or Small Faces. If a face is

  • cut off by the border of the picture; or
  • very small, far in the back

should I still annotate it?

Yes, if it is distinguishable.

- NEW - Certainly faces considerably smaller than the minimum box size can be ignored.

If I Make a Mistake while Correcting an Image. If you moved or deleted a box and then changed your mind, “Restore FaceMatch” will bring back the algorithm’s boxes, overwriting your changes. Consider also “Clear all boxes”, or manipulations of an individual box as explained in Help.

3) “Save” then “Close”, or “Save” then “Next”

When done with an image, hit “Save”, then either:

  • “Close” to return to the Search page.
  • “Next” to navigate to the next picture seen within the current Search-results page. - NEW -

VERY IMPORTANT: When done with an image, always hit “Save”, even if you have made no changes, in order for your review work to be recorded. “Close” or “Next” alone is not enough.

Be sure to save and close your current image before logging-out.

Pausing and Resuming Your Work within a Task, Between Photos

Images that have already been annotated should have a green border; ones not yet done should have a gray border. Resume with the first gray one.

Revisiting a Photo - What Happened to My Changes?

It is not normally necessary to revisit a photo. If you do return to a photo, once again (assuming you are in “show all” mode) you will initially see the FaceMatch annotations; these may take a few moments to appear. To see your earlier annotation work:

  1. Select “show all” mode if not already - NEW -
  2. hit “Clear All Boxes” to remove FaceMatch boxes
  3. hit “Restore”.
Submitting a Completed Task

After you save and close the last image, you must submit the task as completed as follows:

  1. The choice of Export type should be “Final GroundTruthData to TSV file”
  2. Hit the “Export All Results” button, and save the file on your local disk. The filename should automatically reflect the task name and “Final”. If there is any problem with the “Export All Results” button, try hitting the “Search” button up above first to list the images again.
  3. Optional: you can verify the file contents by opening it in a text editor. The expected content is discussed in the next section.
  4. Copy the file name into the Melange “Upload” field, and hit the Upload button.
  5. Change the Melange status for the task to “Ready for Review”.

Expect an acknowledgement within 24 hours, and, after QA, an acceptance within 3 days.

Expected File Format

There will be 1 line for each image visited and “Saved”. Tasks routinely will have 399 images, in which case there should be at least 399 lines in the file.

Each line in the file must contain these fields:

  • an identifier
  • an url
  • a username followed by a filename
  • zero or more annotations of form f[x,y;w,h]

Multiple annotations in a line are separated by tabs. The first letter of an annotation will be “f” for “face” or “p” for profile. Within the brackets, the first two integers will be the coordinates within the photo of the box's upper-left corner, followed by the width and height, all in pixels.

QR Code
QR Code agasti:vesuvius:gci2012:annotation:face (generated for current page)