Exploiting shape cues for weakly supervised semantic segmentation

作者：

Highlights：

• We shed light on the connection between the locality of CAMs and the texture bias of CNNs, which has hardly been handled before.

• We introduce a novel weakly supervised segmentation method that explicitly leverages shape-biased features as shape cues for producing comprehensive segmentation maps, overcoming the locality of CAMs.

• We propose a new pseudo mask generation method, where both color and semantic information are leveraged for obtaining pairwise pixel affinities, thereby accurately refining initial mask predictions.

• On PASCAL VOC 2012, the most popular benchmark, our method achieves a new state-of-the-art performance with a significant margin in both the single and the multi-stage settings.

摘要

•We shed light on the connection between the locality of CAMs and the texture bias of CNNs, which has hardly been handled before.•We introduce a novel weakly supervised segmentation method that explicitly leverages shape-biased features as shape cues for producing comprehensive segmentation maps, overcoming the locality of CAMs.•We propose a new pseudo mask generation method, where both color and semantic information are leveraged for obtaining pairwise pixel affinities, thereby accurately refining initial mask predictions.•On PASCAL VOC 2012, the most popular benchmark, our method achieves a new state-of-the-art performance with a significant margin in both the single and the multi-stage settings.

论文关键词：Semantic segmentation,Weakly supervised learning,Texture biases,Shape cues

论文评审过程：Received 18 December 2021, Revised 14 July 2022, Accepted 1 August 2022, Available online 2 August 2022, Version of Record 6 August 2022.

论文官网地址：https://doi.org/10.1016/j.patcog.2022.108953