Anti-Adversarially Manipulated Attributions for Weakly and Semi-Supervised Semantic Segmentation

Lee, Jungbeom; Kim, Eunji; Yoon, Sungroh

doi:10.1109/cvpr46437.2021.00406

Cited by 193 publications

(156 citation statements)

References 41 publications

Supporting

Mentioning

155

Contrasting

Order By: Relevance

“…2015; Ahn and Kwak, 2018;Lee et al, 2021a). Prevailing WSSS methods with image-level labels usually adopt a multi-step framework.…”

Section: Image Ours Camsmentioning

confidence: 99%

“…Prior works have demonstrated that the first step, i.e. generating initial coarse labels, is crucial to the training of segmentation models and the final segmentation performance (Wang et al, 2020b;Chang et al, 2020b;Lee et al, 2021a). As aforementioned, most methods train classification networks to produce Class Activation Maps (CAMs) (Zhou et al, 2016) as the initial coarse labels.…”

Section: Image Ours Camsmentioning

confidence: 99%

“…Weakly-Supervised Semantic Segmentation (WSSS) aims to develop semantic segmentation models with weak annotations, such as image-level labels (Papandreou et al, 2015;Pinheiro and Collobert, 2015;Ahn and Kwak, 2018;Lee et al, 2021a), points (Bearman et al, 2016), bounding boxes (Song et al, 2019;Oh et al, 2021;Lee et al, 2021b), and scribbles (Lin et al, 2016). In this work, we focus on WSSS with only image-level labels.…”

Section: Wsss With Image-level Labelsmentioning

confidence: 99%

“…proposed to cluster the original semantic categories to sub-categories and further leveraged the sub-category labels to supervise the training of the network, which could enforce the network to discover more object regions to distinguish sub-categories. (Lee et al, 2021a) proposed an anti-adversarial attack approach to gradually pull image away from decision boundaries which helped to discover more object regions. (Zhang et al, 2020b) firstly introduced causal inference (Rubin, 2019) to alleviate the confounding bias attributed by ambiguous boundaries.…”

Section: Improvements On Camsmentioning

confidence: 99%

“…Segmentation Network. For the semantic segmentation network, we used the DeepLabV2 (Chen et al, 2017) system with ResNet101 (He et al, 2016) as backbone, which is a prevailing choice for WSSS task (Ke et al, 2021;Lee et al, 2021a;Chang et al, 2020b). For experiments on PASCAL VOC 2012 dataset (Everingham et al, 2010), we followed the default settings of DeepLabV2 (Chen et al, 2017), i.e., the learning rate was initially set to 0.001 and decayed with a polynomial scheduler.…”

Section: Implementation Detailsmentioning

confidence: 99%

See 4 more Smart Citations

Weakly-Supervised Semantic Segmentation with Visual Words Learning and Hybrid Pooling

Ru¹,

Du²,

Zhan³

et al. 2022

Preprint

View full text Add to dashboard Cite

Weakly-Supervised Semantic Segmentation (WSSS) methods with image-level labels generally train a classification network to generate the Class Activation Maps (CAMs) as the initial coarse segmentation labels. However, current WSSS methods still perform far from satisfactorily because their adopted CAMs 1) typically focus on partial discriminative object regions and 2) usually contain useless background regions. These two problems are attributed to the sole image-level supervision and aggregation of global information when training the classification networks. In this work, we propose the visual words learning module and hybrid pooling approach, and incorporate them in the classification network to mitigate the above problems. In the visual words learning module, we counter the first problem by enforcing the classification network to learn fine-grained visual word labels so that more object extents could be discovered. Specifically, the visual words are learned with a codebook, which could be updated via two proposed strategies, i.e. learning-based strategy and memory-bank strategy. The second drawback of CAMs is alleviated with the proposed hybrid pooling, which incorporates the global average and local discriminative information to simultaneously ensure object completeness and reduce background regions. We

show abstract

“…2015; Ahn and Kwak, 2018;Lee et al, 2021a). Prevailing WSSS methods with image-level labels usually adopt a multi-step framework.…”

Section: Image Ours Camsmentioning

confidence: 99%

Section: Image Ours Camsmentioning

confidence: 99%