Sahana Vesuvius - Tasks for Google Code-in 2012

Annotating Images of Faces with ImageStats [Category: QA/Testing]

 Frontal face example

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.

Skills Needed

Familiarity with using a web browser. Also, the need to judge certain rectangle colors, which may impede those with certain forms of color blindness. No software coding is involved.

Browser Prerequisites

The task requires access to a recent browser that has HTML 5 support. Chrome is recommended. - NEW -

Introduction and Purpose

Research is going on to develop and improve face-detection and face-matching algorithms. Such computer algorithms may ideally help current and future projects involved with preparedness, disaster response, family reunification, and medical assistance, including Sahana Vesuvius and its manifestation as NLM’s People Locator (PL).

To assist with algorithm training/testing and in establishing “ground truth”, you are asked to annotate images of faces with drawn rectangles, or judge such annotations. A web tool, “ImageStats”, has been built for this, using part of the Vesuvius/PL infrastructure and additional open-source components. A customized-for-GCI version is available at:
http://lhce-pl-web01.nlm.nih.gov/imagestats
Over 40,000 public images are available, but the work here will be subdivided into GCI-sized tasks.

A typical task would involve approximately 400 images, each of which has one or more faces. For each face, you must ensure that there is an appropriately-placed rectangle around it, with the rectangle’s color marking it as a front or profile view. Some judgement can be involved in both aspects. Note: at the outset of GCI, during a “pilot phase”, only a limited number of annotation tasks will be available, and these will all be “facial annotation” as just described. After a while, tasks will be more numerous and include “skin annotation” as well.

Claiming a Task

Go to the Melange task list, pick an available ImageStats task, and note the “Event” name there:

  • If it begins with “Face…”, it’s a “Facial Annotation” task;
  • If it begins with “Skin…”, it’s a “Skin Annotation” task.

IMPORTANT: In a Melange comment when claiming your first ImageStats task, include your email address. This will speed up sending you your account username & password.

(Note: Initially, during the pilot phase, tasks were done in duplicate by two different students, indicated by task names beginning in “01_” and “02_”. When that was true, if you had done a particular “01_” task (e.g., “Face 01_05”), then you were asked not to do the “02” task (e.g., “Face 02_05”), or vice versa. However, to simplify things, only the “01_” tasks will be routinely available now.)

Registering Once to Use ImageStats

The ImageStats team needs your email address to register you. Get it to us (but not before you have claimed a task) either by:

  • Including it in a Melange comment as indicated above (recommended);
  • Sending email to imagestatsaccount@nlm.nih.gov, with the subject line “Register for ImageStats Tasks”.

You will be sent back a user name and password, that the mentor will associate with your particular task. Be patient, this is a manual process. Ideally, this response happens before the mentor grants your claim.

Conducting Your Task with ImageStats - Getting your Set of Pictures and Navigating Through Them

Go to http://lhce-pl-web01.nlm.nih.gov/imagestats and log-on with your registered user ID and password.

Remember, any particular task here will be EITHER Facial Annotation, or Skin Annotation, BUT NOT BOTH.

Claiming Another ImageStats Task

Use Melange as before. In a comment, it will be helpful to mention that you already have a username and password. The mentor will then associate your new task with your existing user name, ideally prior to granting your claim.

Questions and Comments

Send email to discuss@sahanafoundation.org

FYI - About the Images

The GCI tasks draw from the “Annotated Facial Landmarks in the Wild” set of over 40,000 public images. As the name implies, there are some annotations available for this set, but these are of a different style than what we need. For instance, the skin annotations collected here will provide a fuller distribution of skin tones under different lighting conditions for training than is currently available. Note that any particular task set of images is not likely to have representative skin tones from the world at large, but the AFLW corpus as a whole plus other planned sources should provide adequate coverage.

FYI - About ImageStats

This was rapid-prototyped from a number of open source components, among them Vesuvius, GWT, GXT, vaadin, Tomcat, mysql, and SOLR. The instance is hosted and managed using NLM’s Vesuvius/PL infrastructure. Current thinking is that ImageStats is overly complicated in structures and dependencies, and might be fruitfully reimplementation in a simpler manner… a possible Google Summer of Code 2013 project. Source code is available here.


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