A Hierarchical-Based Web-Platform for Crowdsourcing Distinguishable Image Patches

Abstract

In this work, we present an open-source web-platform for crowdsourcing a unique kind of image labels. This is done by introducing image segments that we refer to by the term “distinguishable patches”; as the name implies, a distinguishable patch is a small segment of an image that identifies a particular object. Although these distinguishable patches will naturally be part of the object; more often than not, these distinguishable patches combined will not cover the entire body of the object, which makes the nature of the data collected distinct and rather different than what would be obtained from a traditional image segmentation system. To minimize human labeling efforts while maximizing the amount of labeled data collected, we introduce a novel top-bottom hierarchical approach to automatically determine the size and the location of the patches our system will present to individuals (referred to by “workers”) for labeling, based on previously labeled patches. The three processes of: determining the size and location of the patch, assigning a particular patch to the right worker, and the actual cropping of these patches, all happen in real-time and as the workers are actively using our web-platform. As far as the authors are aware, this unique form of image data has not been collected in the past, and its impact has not been explored, which makes this work highly valuable and important to the research community. One of the many ways that the authors are interested in using these distinguishable patches would be to improve the accuracy of a machine learning image classification system, by providing an enriched dataset of images that not only contain a single label for each image, but rather a spatial distribution of the distinguishability of an object in each image.