Analyzing Object-Scale Dependency of Dual-Branch Sigmoid CAM for Weakly Supervised Object Localization

다국어 초록 (Multilingual Abstract)

Class Activation Mapping (CAM) is widely used for interpreting deep visual models and enabling weakly supervised object localization(WSOL), but its reliance on softmax-based classifiers introduces competitive normalization that suppresses weak or co-occurring objectevidence, often resulting in incomplete localization. In this work, we build upon our previously proposed dual-branch sigmoid architectureand conduct a systematic, scale-aware analysis to examine how decoupling localization from softmax normalization affects localizationbehavior across object sizes. Using the ImageNet-1K dataset, we stratify evaluation samples into small-object and large-object subsetsbased on ground-truth bounding box statistics and perform controlled WSOL evaluations across multiple backbones. The results showthat sigmoid-based localization consistently improves localization completeness, with particularly pronounced gains for small objects, whilemaintaining or modestly improving performance on large objects. These findings provide empirical evidence that the limitations ofsoftmax-based CAMs are strongly object-scale dependent and demonstrate that sigmoid-based localization effectively mitigates signalsuppression in scale-sensitive WS OL scenarios, offering a practical framework for diagnosing and improving localization behavior