Computer Vision News - September 2022

23 Razieh Kaviani Baghbaderani by projecting both the source and target data to a so-called abundance space which is regularized by some physical constraints. In this way, the model trained on the source domain would be able to perform well on the target dataset without extra data annotation efforts. The experimental results show stable deployment performance across multiple satellite datasets. datasets for a certain task. This is an inevitable issue in remote sensing area where spectral variability is prevalent due to different acquisition conditions, i.e. illumination and atmospheric conditions. To bridge the gap between different but related datasets, an unmixing-based domain adaptation approach is proposed, as shown in Figure 2. It decreases the domain discrepancy between domains