Crowdsourcing the Acquisition of Video Segmentations by Propagation, Evaluation and Merge
The lack of very large video segmentation datasets is currently an obstacle for creating strong data-driven algorithms for video analysis and rendering (e.g. visual editing for cinema post-production, 2D-3D, etc…) . One of the main reasons of this shortcoming is the lack of scalability in the current approaches used in dataset generation. We propose a cost-efficient and scalable pipeline for large-scale video segmentation through crowdsourcing. We develop methods to incorporate annotations and scribbles of vastly different qualities from crowd workers to generate high-quality segmentations. By leveraging computer vision methods between the tasks for crowd workers and providing the user with effective interfaces for managing the large number of results coming from the crowd, our pipeline can generate high-quality results at a fraction of the cost of a naive distribution of tasks to the crowd workers. We evaluate and discuss each of the design choices we make, and demonstrate the scalability of our method by creating a new video dataset with high-quality annotations.