Weakly supervised object localization (WSOL) models can lose class-discriminative ability under strong domain shifts, often developing a prediction bias toward a small subset of classes and assigning most target samples to those dominant classes. Because source-free unsupervised domain adaptation (SFDA) methods depend on pseudo-labels, such biased predictions lead to poor adaptation. This is especially problematic in computational histopathology, where differences in organ morphology, staining protocols, and acquisition conditions alter the visual cues that models rely on.
The paper introduces SFDA with selective prediction forgetting (SFDA-SPF), which targets this class-imbalance problem without needing the true target class distribution. The method identifies dominant classes from the model's own target predictions and constructs a "forget set" of uncertain samples from those classes near decision boundaries, alongside a "retain set" of confident, accurate samples. A complementary optimization objective reduces the probability assigned to dominant classes for forget-set samples, allowing alternative class assignments to emerge while retain-set samples anchor reliable predictions.
SFDA-SPF iteratively reassesses which target images belong to the forget and retain sets, and it jointly optimizes the pixel-level classifier alongside image-level pseudo-label estimation. This joint optimization is intended to improve localization, not just classification. The source is a single arXiv paper, so the results are presented as the authors' proposal rather than an independently verified finding.