Efficient Annotation of Segmentation Datasets
with Polygon-RNN++

David Acuna* 1,3
Huan Ling* 1,2
Amlan Kar* 1,2
Sanja Fidler1,2

1University of Toronto
2Vector Institute
CVPR, 2018

We learn to predict interactive polygonal annotations of objects to make human annotation of segmentation datasets much faster.

Manually labeling datasets with object masks is extremely time consuming. In this work, we follow the idea of PolygonRNN to produce polygonal annotations of objects interactively using humans-in-the-loop. We introduce several important improvements to the model: 1) we design a new CNN encoder architecture, 2) show how to effectively train the model with Reinforcement Learning, and 3) significantly increase the output resolution using a Graph Neural Network, allowing the model to accurately annotate highresolution objects in images. Extensive evaluation on the Cityscapes dataset shows that our model, which we refer to as Polygon-RNN++, significantly outperforms the original model in both automatic (10% absolute and 16% relative improvement in mean IoU) and interactive modes (requiring 50% fewer clicks by annotators). We further analyze the cross-domain scenario in which our model is trained on one dataset, and used out of the box on datasets from varying domains. The results show that Polygon-RNN++ exhibits powerful generalization capabilities, achieving significant improvements over existing pixel-wise methods. Using simple online fine-tuning we further achieve a high reduction in annotation time for new datasets, moving a step closer towards an interactive annotation tool to be used in practice.

# work done when D.A. was at UofT



David Acuna*, Huan Ling*, Amlan Kar*, Sanja Fidler

Efficient Annotation of Segmentation Datasets
with PolygonRNN++

CVPR, 2018. (to appear)

[Online Demo]
[PolyRNN (CVPR '17]]
For commerical enquiries, please contact Sanja Fidler (fidler@cs.toronto.edu)


 [Inference]  [Training/Tool]


Automatic annotation on Cityscapes (given GT bounding boxes)
Automatic annotation on ADE20k (given GT bounding boxes, using model only trained on Cityscapes). Ours in Yellow, GT in white

Average number of clicks per instance required in our model (left and up is better)

We gratefully acknowledge NVIDIA corporation for the donation of GPUs used for this research. This webpage template was borrowed from Richard Zhang.