Computer Vision News - June 2018

T_2 is the IoU between the predicted polygon and the GT mask, with polygons achieving agreement above T_2 considered to need no interference from the annotator. In the results above are the average number of clicks per instance required to annotate all classes on the Cityscapes val set with different values of T_2. At T_2 = 0.8 the new model is still more accurate than Polygon-RNN at T_2 = 1.0. At T_2 = 0.7, it achieves over 80% IoU with an average of 5 clicks per object, over 50% reduction. Conclusions: This paper proposed Polygon-RNN++, a model for object instance segmentation that can be used to interactively annotate segmentation datasets. The model builds on Polygon-RNN presented in CVPR-2017, while introducing important improvements to outperform the previous approach, in both automatic and interactive modes. For the interested reader: in the original article the authors show (a) the model’s robustness to noisy annotators and (b) a capability to generalize to novel domains. Research 9 Research Computer Vision News