Combining crowd worker, algorithm, and expert efforts to find boundaries of objects in images

Abstract

While traditional approaches to image analysis have typically relied upon either manual annotation by experts or purely-algorithmic approaches, the rise of crowdsourcing now provides a new source of human labor to create training data or perform computations at run-time. Given this richer design space, how should we utilize algorithms, crowds, and experts to better annotate images? To answer this question for the important task of finding the boundaries of objects or regions in images, I focus on image segmentation, an important precursor to solving a variety of fundamental image analysis problems, including recognition, classification, tracking, registration, retrieval, and 3D visualization. The first part of the work includes a detailed analysis of the relative strengths and weaknesses of three different approaches to demarcate object boundaries in images: by experts, by crowdsourced laymen, and by automated computer vision algorithms. The second part of the work describes three hybrid system designs that integrate computer vision algorithms and crowdsourced laymen to demarcate boundaries in images. Experiments revealed that hybrid system designs yielded more accurate results than relying on algorithms or crowd workers alone and could yield segmentations that are indistinguishable from those created by biomedical experts. To encourage community-wide effort to continue working on developing methods and systems for image-based studies which can have real and measurable impact that benefit society at large, datasets and code are publicly-shared (http://www.cs.bu.edu/~betke/BiomedicalImageSegmentation/).