Weakly-supervised semantic segmentation via online pseudo-mask correcting

Abstract

Many existing weakly-supervised semantic segmentation methods focus on obtaining more accurate pseudo-masks with weak labels. So far pseudo-masks have come close to the ground truth. However, the potential of these high-quality pseudo-masks has not been fully explored. This is because pseudo-masks inevitably contain partial noisy labels. Deep segmentation networks tend to overfit noisy labels, which leads to poor generalization performance. In this work, we propose a new method to mitigate the damage caused by noisy labels. First, We use the exponential moving average (EMA) model of the online segmentation model as the teacher. Then, predictions from the teacher model are used to correct pseudo-masks online. Besides, learning with noisy labels has been extensively studied in classification tasks. We also introduce these anti-noise techniques and find them also effective for the segmentation task. Our proposed method can be easily embedded into existing weakly-supervised semantic segmentation algorithms and bring 2.3% IoU improvement without expensive computational cost. It also achieves the state-of-the-art performance on the PASCAL VOC 2012 benchmark dataset.