| Summary: | Semantic segmentation has recently emerged as a prominent area of interest in Earth observation. Several semantic segmentation datasets already exist, facilitating comparisons among different methods in complex urban scenes. However, most open high-resolution urban datasets are geographically skewed toward Europe and North America, while coverage of Southeast Asia is very limited. The considerable variation in city designs worldwide presents an obstacle to the applicability of computer vision models, especially when the training dataset lacks significant diversity. On the other hand, naively applying computationally expensive models leads to inefficacies and sometimes poor performance. To tackle the lack of data diversity, we introduce a new UAVPal dataset of complex urban scenes from the city of Bhopal, India. We complement this by introducing a novel dense predictor head and demonstrate that a well-designed head can efficiently take advantage of the multiscale features to enhance the benefits of a strong feature extractor backbone. We design our segmentation head to learn the importance of features at various scales for each individual class and refine the final dense prediction accordingly. We tested our proposed head with a state-of-the-art backbone on multiple UAV datasets and a high-resolution satellite image dataset for LULC classification. We observed improved intersection over union (IoU) in various classes and up to 2<inline-formula><tex-math notation="LaTeX">$\%$</tex-math></inline-formula> better mean IoU. Apart from the performance improvements, we also observed nearly 50<inline-formula><tex-math notation="LaTeX">$\%$</tex-math></inline-formula> reduction in computing operations required when using the proposed head compared to the traditional segmentation head.
|
|---|