Підвищення точності локалізації в навчанні зі слабким контролем за допомогою базових моделей і уточнення псевдоміток з урахуванням невизначеності

Abstract

У цьому дослідженні запропоновано нову модельну структуру для вирішення проблеми низької точності локалізації в задачах розпізнавання та сегментації об'єктів із використанням слабоконтрольованого навчання.

This paper considers a novel framework to address the persistent challenge of low localization precision in weakly supervised learning for object detection and segmentation tasks. Despite recent advancements, existing WSL methods often struggle to accurately delineate object boundaries due to reliance on ambiguous and noisy supervisory signals [1]. The approach presented in this study integrates a vision foundation model – the Segment Anything Model (SAM) – with uncertainty-guided pseudo-label refinement to enhance the quality of weak supervision. Specifically, SAM is employed to generate initial segmentation masks from minimal image-level annotations, providing a strong prior on object locations. To further mitigate the impact of noisy predictions, an uncertainty estimation mechanism filters out lowconfidence pseudo-labels, ensuring that only reliable supervision guides model training [2]. Experimental results on standard benchmark datasets demonstrate that the proposed method significantly improves localization accuracy compared to state-of-the-art WSL approaches, while simultaneously reducing annotation costs. This work highlights the potential of combining foundation models with uncertainty-aware learning strategies to bridge the performance gap between weakly and fully supervised object localization.