By Akshay Asija
Facebook is in a difficult situation. While all of the companys social media products are recording a steady user and revenue growth, its public image has been dented. The biggest blow to Facebooks reputation was dealt in 2016, when its Newsfeed algorithms were manipulated by (mainly) Russian agents and exploiters to influence the way people voted in the US presidential election.
Developments at Facebook
On several occasions in the past year, Facebooks PR executives and CEO Mark Zuckerberg have apologised for its missteps. The social media giant has rolled out a few initiatives to tackle the issues that have been plaguing the site for some time. The company is using machine learning and intelligent algorithms to identify clickbait, trolls and fake news and remove these from the users feeds. Facebook has also made some fundamental changes to the way content is presented to its users so that it is more people-oriented. Thus, after years of making Facebook a lucrative platform for advertisers (and almost falling out of favour of its user base), the company is trying to emphasise the fact that it is a platform built primarily with users in mind. The Silicon Valley giant isnt just driving this message to its users, but also to the developer community.
Facebook has an extensive research division that works on solving problems in diverse fields such as machine learning, economics and networking, among others. This department is probably the reason why the company often poaches employees from rival technology companies like Google, Apple, and Amazon. A large number of academicians and researchers also work at Facebook to improve its underlying code and make the service better at analysing user data.
Computer Vision
One of the core research areas at Facebook is computer vision, which is a nascent discipline in the broader field of artificial intelligence. It involves processing and analysing of images and videos to obtain high-level inferences. Computer vision tasks, which are devised using statistical, mathematical and geometrical models, scrutinise the visual data stored in images and convert it into numerical and symbolical form. This data can then be processed by a computer to make appropriate decisions. An ideal computer vision-based device would be able to replicate the way the human visual system (the part of the nervous system that allows us to process visual details) functions.
The primary concept involved in computer vision is object detection. Object detection involves locating specific objects, such as humans, vehicles, etc. in digital imagery. Face detection, which forms the basis of a lot of smartphone photography today, is a highly specific application of object detection. It is also used in image retrieval, security, surveillance, and advanced driver assistance systems (ADAS). The procedure for carrying out object detection varies from object to object, depending on the structure of the object. Most object detection algorithms, however, employ machine learning and extracted features to identify the instances of a given category of objects. All object detection algorithms are typical implementations of one of the following models: deep learning object detection (which use the technique described above), feature-based object detection, Viola-Jones object detection (which uses a summation technique on image pixels, similar to mathematical Haar functions), histogram of oriented gradients technique (which counts occurrences of gradient orientation in an image), and blob analysis (which often involves the removal of an images background for a better look at the remaining objects called blobs). Some other, less popular approaches to object detection are gradient-based, derivative-based, and template matching methods.
The scientists working on computer vision systems at Facebook are building visual sensors derived from digital images and videos that can study users environment, and let Facebook services automate certain tasks that people automatically do today visually. Facebook aims to use its work on computer vision to tailor its services to a user, based on their environment. Object detection has been a prominent subject of research at the company, and the company even created a dedicated platform for the same, Detectron.
Detectron, now open source
In a move that is likely to please developers building computer vision-based software, Facebook recently announced that it is open sourcing Detectron. The social media behemoth has long been a proponent of the open source movement. It has released a significant amount of its code to the public for understanding, improvement and usage. While many consider Facebook a company that is hostile to the needs of its users, the sentiment is pretty much the opposite when it comes to software developers. Facebook is routinely praised by software developers owing to the importance it gives to emerging technologies and the fact that it encourages community-based development.
What makes it special
The Detectron project began in July 2016 at Facebook AI Research (FAIR). It was intended to serve as an efficient object detection system built on the Caffe2 deep learning framework. The platform was extensively used in many of the companys research projects, including Mask R-CNN, which won the Marr Prize at International Conference on Computer Vision (ICCV) last year. The algorithms described in the projects that use Detectron provide intuitive models for important computer vision tasks, and have immensely helped development in visual perception systems in the past few years.
What makes Detectron special is that it is extremely flexible, and hence, preferred over other object detection systems for the purposes of development. It is also very fast and can be real-time tasks. Facebooks employees have used this platform for training their own models for multiple applications, such as augmented reality. Detectrons reliance on Caffe2 allows for a relatively convenient deployment of trained models in the cloud and on mobile devices.
How is Detectron going to change things?
Facebook says that its intention behind open sourcing Detectron is to make its research as open as possible and to aid research all around the world. The company hopes that the research community will replicate the results it has obtained and have access to the resources that FAIR has.
Given the number of projects that use this relatively new platform (FAIR, on its website, lists ten projects that rely on Detectron), it is not hard to see that Detectron can be immensely useful. Going forward, we can see improved versions of facial recognition and augmented reality software, powered by this newly open-sourced platform. Detectron might revolutionise the domains of AR and computer vision, the way React (Facebooks open source JavaScript library for building user interfaces) did for web and app development.
Featured Image Source: Visualhunt