Poor's man Rekognition

Amazon Rekognition is a (paid) service that is able to identify objects, people, text, scenes, and activities in a picture. We want to produce a free alternative.

While being able to do everything Rekognition does over the course of a summer is unrealistic, we think we should be able to kickstart the effort and get to a point where the project will be usable and attract more developers to the effort.

Let's start with faces, and this is the actual goal: Given a set of properly tagged people (suppose for example, a number of celebrities), create an API that can be used to identify such people in other images.

At a minimum

- Your code must be able to run either locally or in a cloud service - The API needs to be callable from any language (so REST, or something similar) - You need to provide a native binding to a language of your choice - You will need to provide the sample images yourself (can be taken from the internet) - For any picture, it will identify the faces there, and for each face give a list of the most likely people from the known set in order of likeness - Must be able to learn from user feedback

Video processing

Building on the previous work, figure out a way to analyze a video and determine who is in each scene. For example if you detect a known actor in let's say, 3:05, and the scene runs until 3:17, then that actor is there from 3:05 to 3:17. Generate a text file (any format) that lists who is in each scene. If possible, don't brute force.

Notes

- You can use any open source library to help with the project as long as it doesn't prevent from meeting the full scope

How to get started

Since this is a new project we don't have issues open on it. A good way to start, and what other students are doing, is to write a working proof of concept that shows you are capable to doing the full thing. And then come up with ideas and a plan to implement them during the summer.

Mentor

Johannes Von Lochter <johannes.lochter at facens d-o-t br> (look him up)