ScribbleSup: Scribble-Supervised Convolutional Networks for Semantic Segmentation

ScribbleSup: Scribble-Supervised Convolutional Networks for Semantic Segmentation

18 Apr 2016 | Di Lin1*, Jifeng Dai2 Jiaya Jia1 Kaiming He2 Jian Sun2
The paper "ScribbleSup: Scribble-Supervised Convolutional Networks for Semantic Segmentation" by Di Lin, Jifeng Dai, Jiaya Jia, Kaiming He, and Jian Sun proposes a method to train convolutional networks for semantic segmentation using scribbles as annotations. The authors note that while large-scale data is crucial for improving semantic segmentation models, annotating per-pixel masks is time-consuming and inefficient. Scribbles, a user-friendly method for interactive image segmentation, are widely used and can be annotated more quickly and easily than precise masks. The proposed algorithm uses a graphical model to propagate information from scribbles to unmarked pixels and to learn network parameters. The method is evaluated on the PASCAL VOC dataset, showing competitive results compared to strongly-supervised methods. Additionally, the method is effective on the PASCAL-CONTEXT dataset, which involves objects and stuff that are often difficult to annotate with precise masks. The authors demonstrate that scribble annotations can be a cost-effective solution for increasing the accuracy of semantic segmentation models.The paper "ScribbleSup: Scribble-Supervised Convolutional Networks for Semantic Segmentation" by Di Lin, Jifeng Dai, Jiaya Jia, Kaiming He, and Jian Sun proposes a method to train convolutional networks for semantic segmentation using scribbles as annotations. The authors note that while large-scale data is crucial for improving semantic segmentation models, annotating per-pixel masks is time-consuming and inefficient. Scribbles, a user-friendly method for interactive image segmentation, are widely used and can be annotated more quickly and easily than precise masks. The proposed algorithm uses a graphical model to propagate information from scribbles to unmarked pixels and to learn network parameters. The method is evaluated on the PASCAL VOC dataset, showing competitive results compared to strongly-supervised methods. Additionally, the method is effective on the PASCAL-CONTEXT dataset, which involves objects and stuff that are often difficult to annotate with precise masks. The authors demonstrate that scribble annotations can be a cost-effective solution for increasing the accuracy of semantic segmentation models.
Reach us at info@study.space
[slides] ScribbleSup%3A Scribble-Supervised Convolutional Networks for Semantic Segmentation | StudySpace