Atme Project with Kobe University
Our team VIV has collaborated with Ariki Professor’s research team in Kobe University in Japan to develop an original face recognition system.
The key to success of machine learning is a scalable database. We create an open platform and ask contributors to play around with this simple app. We are still keeping improving the website and hope to meet the target collecting 2.8 million face photos, which is not a small number. I am the person in charge of the maintenance of photo collection system. We first created a platform to collect photos of faces and categorize them into simple groups and open to the public to help with assigning them to relevant emotions.
The COMM.UNITY platform serves four functions:
- Collection of photos from different sources e.g. Flickr, Tumblr with their API
- Push all the collected photos to Face++ to analyze basic features and crop the facial features. ( Face++ is an open source Face recognition platform which is able is measure the level of smile, facial features, but not other emotions)
- Let participants match photos of faces with relative emotions with an intriguing user interface
- An audit system for a final screening of photos to maximizing accuracy
At the pre-launch stage, I designed the landing page and UI. For example, I changed the buttons to assign seven emotions into Emoji.
After the building up of the system, I took up the role as the project manager. I am also in charge of system maintenance and planning of improvement. I have been adding new photo collection mechanisms with different new sources and APIs. Flickr becomes the major source of the photo, ( Instagram API is not working anymore for searching photos by tag), while functions to collect photos from Google Search, Tumblr and Twitter were also added consecutively. In addition, a new feature was added to audit the screen pictures. Improvement of the loading speed is also necessary as the photos and checking record accumulated. Currently, in order to build the deep learning system in Babylook, our team is adding functions of a collection of photos of baby.
System development with Laravel
The image below shows one function I implemented for uploading photos which do not have source URI and its integration to the existing system.
It illustrates how the pictures uploaded by Direct Uploader are cropped, as there are no URI of the source images, while for photos retrieved from Flickr, we could get the URI of the image source and cropping can be done from using the source photos. The photos were retrieved from the S3 server and then processed by Face++ for its first time to detect the existence and number of faces presented and its second time to Face++ to crop individual faces.
You can also help!
Help the world unite together and break the language barrier. For this, we need your help to detect faces that we have collected. Select the matching expression for the face photos. Register an account on the COMM.UNITY website and start checking photos!