The increased availability of high resolution satellite imagery allows tosense very detailed structures on the surface of our planet. Access to suchinformation opens up new directions in the analysis of remote sensing imagery.However, at the same time this raises a set of new challenges for existingpixel-based prediction methods, such as semantic segmentation approaches. Whiledeep neural networks have achieved significant advances in the semanticsegmentation of high resolution images in the past, most of the existingapproaches tend to produce predictions with poor boundaries. In this paper, weaddress the problem of preserving semantic segmentation boundaries in highresolution satellite imagery by introducing a new cascaded multi-task loss. Weevaluate our approach on Inria Aerial Image Labeling Dataset which containslarge-scale and high resolution images. Our results show that we are able tooutperform state-of-the-art methods by 8.3% without any additionalpost-processing step.
展开▼
