Unsupervised Domain Adaptation with Debiased Contrastive Learning and
Support-Set Guided Pseudo Labeling for Remote Sensing Images
Abstract
Here, we address the challenge of generalizing object detection and
labeling in remote sensing data from one dataset to another. It is the
variability in different altitudes, geographical variances, weather
conditions, and object size across datasets that state-of-the-art DNN
largely fails to generalize. Contrastive-based unsupervised domain
adaptation attempts to bridge the gap by producing discriminating
features for frames and instances across different datasets, and we have
shown some progress on the domain alignment on local, global, and
instance-aware features for remote sensing data. In this research, we
propose using support-guided pseudo-labeling on the target domain
instances to enable instance domain adaptation on the source and target
datasets. Next, we introduce the contrastive loss function with multiple
positive examples to make the model more generalized of the variable
appearance of a particular class over images and domains. Also, we
proposed debiased contrastive learning based on class probabilities to
address the challenge of false negatives in the unsupervised framework.
We show the advantages of the proposed model on satellite (DIOR and
DOTA2.0) and drone (Visdrone and UAVDT) image datasets.