Incorporating pixel proximity into answer aggregation for crowdsourced image segmentation

The success of crowdsouring has been witnessed in handling a wide spectrum of application tasks. However, there remain technical challenges for dealing with complex tasks like image segmentation. One reason is that macro tasks usually have complex hidden internal structures, which are difficult to be captured and utilized. In this work, we are concerned with answer aggregation of crowdsourced image segmentation. Recent crowdsouring research on both general tasks and image segmentation ignores the hidden structure information inside a task. To fill the gap, we propose a method named CrowdSeg to aggregate the set of segmentations given by crowdsourcing workers so as to obtain satisfying image segmentation results. First, we propose a model based on a convolutional auto-encoder that can extract the proximity information of adjacent pixels and represent it as embedding features. Second, since each pixel could be background or object, we then cluster the pixels based on embedding features with k-means algorithm and determine whether the cluster is the object or background according to the result accepted by the majority. The proposed method outperforms the baselines on four real-world datasets of biomedical images, which are collected from workers of the real-world crowdsourcing platform. The experimental result shows that our method is more effective and stable than other methods. We have released the real-world datasets and source code for all experiments.