| In the field of computer vision,semantic segmentation has always been a highly significant branch.Under the upsurge of deep learning,semantic segmentation has also developed tremendously.However,the current segmentation networks that have achieved excellent results all require pixel-level semantic labels.In practical situations,obtaining pixel-level annotations requires high costs.Therefore,more and more researchers began to study the use of weakly supervised label information for image semantic segmentation tasks.The research direction of this topic is based on the semantic segmentation task under the image category label.Most of the recently proposed methods rely on object attention to provide supervision cues for training the segmentation network.Thus,the quality of the object attention becomes a crucial bottleneck for improving the final segmentation accuracy.However,since the global average pooling operation is used in the classification network,the nature of the operation itself assumes that the feature vectors at different positions contribute equally to the global representation,ignoring the difference in the contribution of local features to the global representation.It will lead to inaccuracy of the obtained salient regions,which will affect the final result.In this paper,we propose a light weight adaptive pooling(AP)module that helps the classification network generate high quality object attention by adaptively rescaling the contributing weights of different locations in the high-level feature map for categorical prediction throughout the training of image classification.We empirically show the effectiveness of our approach by evaluating it on the challenging PASCAL VOC 2012 semantic segmentation dataset.The experimental results demonstrate that despite its simplicity,our AP module significantly improves the quality of the object attention and incurs much lower training cost compared with other competitive baseline methods under the same conditions. |
PDF Full Text Request