Tackling Background Misclassification in Box-Supervised Segmentation: A Background Constraint Approach

In contrast to the well-established paradigm of supervised instance segmentation, the emerging weakly supervised methodology encounters numerous challenges. One significant difficulty lies in the absence of pixel-wise labels, making it challenging for the model to learn a mapping from the input image to instance masks. For box-supervised instance segmentation, previous method encouraged similar neighboring pixels to share identical labels, yielding reasonably successful outcomes. Yet, such method does not perform well when differentiating highly similar background and foreground pixels, often misclassifying background as foreground due to methodological limitations. To address this problem, our proposed Background Constraint Approach (BCA) introduces a background-specific loss function to reduce misclassification of background pixels, while a complementary foreground loss function mitigates potential negative effects on foreground pixel classification. Our approach has outperformed the baseline by improving the Average Precision (AP) by 1.09% on the COCO val 2017 dataset. Visualization of the inference images confirms that our method more accurately classifies pixels previously mislabeled as foreground by the baseline.