Polygon-RNN is an annotation tool for labeling object instances with polygons. It assumes that the user provides the bounding box around the object. The method then predicts a (closed) polygon outlining the object using a Recurrent Neural Network. It allows the user to correct a predicted vertex of the polygon at any time step if needed, which will be integrated in the prediction task.

Authors claim that this approach speeds up the annotation process by a factor of 4.7 across all classes in Cityscapes dataset, while achieving 78.4% agreement in IoU with original ground-truth, matching the typical agreement between human annotators.

An improved version of this tool is called Polygon-RNN++.

Related publications:

  • Lluis Castrejon, Kaustav Kundu, Raquel Urtasun, Sanja Fidler, "Annotating Object Instances with a Polygon-RNN", IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2017.

Related group: