Weakly-Supervised Salient Object Detection through Object Segmentation Guided by Scribble Annotations

EasyChair Preprint 6770

Abstract

With the advent of Neural Network, Fully-supervised salient object detection achieves great success. However, it takes plenty of efforts to obtain precise pixel-level annotations. In order to reduce human labeling efforts, some research adapt weak form annotations, but they still fall short of the fully-supervised. In this paper, we propose a novel  weakly-supervised salient object  detection framework, which can reduce labeling efforts by using scribble  annotations. In the meantime, we also incorporate Deep Convolutional Network to achieve high performance. To this end, we utilize high-quality region hierarchies, which are generated by Convolutional Oriented Boundary (COB) network, to select optimal level for object segmentation. We build initial saliency maps and thoroughly annotate the images during the initialization phase by spreading labels information from scribbles to other regions. During the training  phase, the salient object detection convolutional network is trained  using the initial saliency maps. Then, we utilize Conditional Random Field (CRF) to refine saliency maps, which will then be used to retrain the network. To  achieve quality saliency maps, we iteratively optimize the  training process. Extensive experiments on six benchmarks demonstrate that our proposed method outperforms previous weakly-supervised algorithms.

Keyphrases: Salient object detection, Scribble annotations, hierarchical segmentation, weakly supervised

@booklet{EasyChair:6770,
  author    = {Xiongying Wang and Zaid Al-Huda and Bo Peng},
  title     = {Weakly-Supervised Salient Object Detection through Object Segmentation Guided by Scribble Annotations},
  howpublished = {EasyChair Preprint 6770},
  year      = {EasyChair, 2021}}