.png){kind=icon}
Return
Rethinking erasing strategy on weakly supervised object localization
.png){kind=icon}
PRE
AI
DOI：10.1016/j.image.2025.117280.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Weakly supervised object localization (WSOL) is a challenging task that aims to locate object regions in images using image-level labels as supervision. Early research utilized erasing strategy to expand the localization regions. However, those methods usually adopt a fix threshold resulting in over- or under-fitting of the object region. Additionally, recent pseudo-label paradigm decouples the classification and localization tasks, causing confusion between foreground and background regions. In this paper, we propose the Soft-Erasing (SoE) method for Weakly Supervised Object Localization (WSOL). It includes two key modules: the Adaptive Erasing (AE) and Flip Erasing (FE). The AE module dynamically adjusts the erasing threshold using the object's structural information, while the noise information module ensures the classifier focuses on the foreground region. The FE module effectively decouples object and background information by using normalization and inversion techniques. Additionally, we introduce activation loss and reverse loss to strengthen semantic consistency in foreground regions. Experiments on public datasets demonstrate that our SoE framework significantly improves localization accuracy, achieving 70.86% on GT-Known Loc for ILSVRC and 95.84% for CUB-200-2011.
Keywords:
Weakly supervised learning
Erasing strategy
Object localization
Journal
S
IF:
2.7
Papers:
2.8K
Citations:
4.2K
Organization
Cited Papers
ShareSaveShareSaveShareSaveShareSaveShareSaveAge and Date for Early Arrival of the Acheulian in Europe (Barranc de la Boella, la Canonja, Spain).png)
PLoS ONE
OAAI
PLoS ONEIF0
OAAIShareSave
View more