SCIDA: Self-Correction Integrated Domain Adaptation From Single- to Multi-Label Aerial Images

Most publicly available datasets for image classification are with single labels, while images inherently multi-labeled in our daily life. Such an annotation gap makes many pre-trained single-label models fail practical scenarios. This issue is more concerned aerial images: Aerial data collected from sensors naturally cover a relatively large land area multiple annotated datasets, which (e.g., UCM, AID), single-labeled. As manually annotating multi-label would be time/labor-consuming, we propose novel self-correction integrated domain adaptation (SCIDA) method automatic learning. SCIDA weakly supervised, i.e., automatically learning the model using massive, images. To achieve this goal, Label-Wise self-Correction (LWC) module to better explore underlying label correlations. also unsupervised (UDA) single- possible. For training, proposed only uses information yet requires no prior knowledge of data; and it predicts labels In experiments, trained single-labeled MAI-AID-s MAI-UCM-s tested directly on Multi-scene Image (MAI) dataset.