Game-Based Crowd Sourcing for Object Labeling

Abstract

Object detection and recognition is one of the primary tasks in computer vision. Labels identify an object detected from an image and help towards further analysis. One of the challenges that Big data caused by the internet era is that with velocity of data there is variety and veracity. For machine learning models to succeed, domain specific datasets are important. Crowd sourcing data is an approach that has been around for almost over a decade now. It allows for more human reach. Game-based crowd sourcing keeps people engaged as they contribute to research and development. Many web game-based crowd sourcing platforms exist for different domains. ihearu-play [1] is one of these made for speech analysis. CloudCV [2] is another made for image analysis. This project is a web platform that functions as game-based crowd sourcing a for object labeling on indoor images. The outcome shall be a dataset for indoor images that can be utilized for research or development purposes. The target is to get users or players to choose from a set of labels, an appropriate label for the object highlighted from an image. The user will be posed to be playing against the computer itself which derives the computer versus human concept. Object detection is performed in python by a pretrained model on Faster RCNN [3] using GluonCV [4] package with MxNet [5] package back-end. For storage between the python model and web pages, MongoDB is used by PyMongo [6] python library and visualized by Robo3T. The web pages are made utilizing HTML canvas, HTML5, CSS3 and Javascript. The pages are first mapped and wireframed on Adobe XD tool. They are later coded and implemented. The images are divided into levels to encourage user engagement. This project is a contribution to the research are of computer vision and further contributions will be made in this research area.