UAVid++: Higher-Quality Labels and Expanded Semantic Taxonomy for Aerial Semantic Segmentation" has been accepted to IEEE Transactions on Geoscience and Remote Sensing (TGRS).

Happy to share that our paper, "UAVid++: Higher-Quality Labels and Expanded Semantic Taxonomy for Aerial Semantic Segmentation" has been accepted to IEEE Transactions on Geoscience and Remote Sensing (TGRS). UAVid++ is an improved and expanded version of the UAVid (https://uavid.nl/) dataset for aerial semantic segmentation. We manually corrected the labels in all the annotated frames and added three new classes (Water, Sky, and Roof), bringing the taxonomy from 8 to 11 categories. We re-trained and re-evaluated five state-of-the-art methods on the dataset variants. The corrected labels alone improve performance by up to +7.4% mIoU, and the new classes add another +4.1%. The benchmark also becomes more discriminative and changes how the SOTA methods rank. We also adapted a frozen DINO-pretrained ViT backbone with a lightweight trainable head, reaching 82.04% mIoU on UAVid++ (+2.7% over the best SOTA method), while outperforming the state of the art on multiple out-of-distribution datasets. The project page, code, and dataset and are all public if you'd like to explore or build on them: Project page: https://lnkd.in/d_vHxybs Code: https://lnkd.in/d3C9yuN2 Dataset: https://lnkd.in/dGWYtDei