Topcoder Age & Gender Determination Facial Analysis Android App - Accuracy Improvement Challenge

Welcome to the Topcoder Age & Gender Determination Facial Analysis Algorithm - Accuracy Improvement Challenge

Challenge Requirements

We ran a proof-of-concept challenge a few weeks back to develop a PoC algorithm. The purpose of this challenge is to increase the accuracy of the algorithm so that it has 80% or higher accuracy in detecting a users age & gender.

Please use the winning submission from the previous challenge (provided in the contest forum) and improve its accuracy.

The available dataset includes 50,000 records of a general population and has multiple variables associated with age and gender. The dataset is semi-structured.

 

Success Criteria

For this challenge, the submissions must implement an algo that meet the following requirements:

• determine the age and gender of a person based on a photo

• must be 80% accurate or higher

-the age can be within a span of 5 years, for example it could suggest a person is between the ages of 30-35 (rather than an exact figure)

-it SHOULD be able to do this with multiple faces if there are more than one face detected

-There must be no crashes in the application

-Please make sure that your source code is well modularised and well commented

Dataset

Sample dataset & images are available below. This dataset will be used for reviewing the submissions.

• https://data.vision.ee.ethz.ch/cvl/rrothe/imdb-wiki/

 

Resources

The following reference links will be helpful

• https://data.vision.ee.ethz.ch/cvl/rrothe/imdb-wiki/

• http://www.robots.ox.ac.uk/~vgg/software/vgg_face/

• https://cmusatyalab.github.io/openface/

• https://bytefish.de/blog/gender_classification/

 

Consideration for Future Use

• Please note that we will eventually use this algorithm to make these detections locally on an Android device using the live camera stream, and that no images will be saved to the device.

• It will be a client side Android app that works offline so no web services can be used for determining age or gender.

 

Review Criteria

• The submissions will be judged by the subjectively on a scale of 1-10 based on the following criteria & the provided dataset. A score of 5 is the minimum passing score and any submissions that do not meet the provided requirements will be failed

  • Adherence to mentioned requirements

  • Accuracy (80% or more)

  • Code quality

Final Submission Guidelines

• Algorithm implementation source code - the solution must be implemented in Java so that it can be incorporated into the Android app.

• Detailed description of Algorithm including all references, links

• Demo video of your solution in action