A Comparative Study of CAM Variants with GrabCut Refinement for Large-Scale Pseudo-Mask Generation

Semantic segmentation typically demands laborious pixel-level annotations, motivating cost-effective weakly supervised approaches that generate pseudo-masks from classifier signals. Weakly supervised pseudo-mask generation offers a cost-effective alternative to costly pixel-level annotation by leveraging classifier activation maps and simple refinement. In this work, we present the first large-scale, systematic comparison of six Class Activation Map (CAM) variantsCAM, Grad-CAM, Grad-CAM++, Eigen-CAM, Layer-CAM, and Ablation-CAMwithin a unified GrabCut refinement pipeline. For each CAM method, we binarize heatmaps using both a fixed 0.2 threshold and Otsus method, then refine seeds with GrabCut to produce final masks. Evaluated on 23,151 ImageNet-LID validation images, our results show that (1) vanilla CAM yields the most precise seeds and highest final mask quality (IoU = 0.5124, Dice = 0.6219), (2) GrabCut consistently improves mask accuracy across all variants (IoU0.10-0.17,Dice0.08-0.14), and (3) although newer CAM methods exhibit superior localization in raw heatmaps, those advantages do not translate into better thresholded pseudomasks after GrabCut refinement. These findings guide practitioners toward simple yet effective weakly supervised segmentation pipelines and highlight directionssuch as multithreshold schemes or alternative refinement methodsfor capturing the strengths of advanced CAM variants.