Data for exploring explainable AI, contrastive and adversarial methods for weakly supervised semantic segmentation

Specifically, it includes: (i) pseudo segmentation masks produced by the P-NOC framework; (ii) pseudo saliency proposals generated by the C²AM-H method; and (iii) segmentation predictions obtained during the verification stage using a DeepLabV3+ model trained with the corresponding pseudo segmentation masks for both Pascal VOC 2012 and MS COCO 2014 datasets.

In addition, the same categories of data are provided for the CSRM framework, including its pseudo-labels and resulting segmentation proposals. These artifacts are released to support reproducibility, qualitative inspection, and further analysis of the weakly supervised learning pipelines investigated in this work.